Recent studies indicate that effectively utilizing inference-time compute is crucial for attaining good performance from large language models (LLMs). Specifically, the Best-of-N (BoN) inference strategy, where an LLM generates multiple responses and a verifier selects the best, has shown strong empirical performance. Motivated by this, we develop a novel inference-aware fine-tuning paradigm, which encompasses the BoN-aware inference framework as a special case. We devise the first imitation learning and reinforcement learning (RL) methods for fine-tuning LLMs using BoN, overcoming the challenging, non-differentiable argmax operator in BoN. We empirically demonstrate that our BoN-aware models implicitly learn a per-example "meta-strategy", which interleaves best responses with more diverse responses that might be better suited to a test-time input—a process reminiscent of the exploration-exploitation trade-off in RL. Our experiments demonstrate the effectiveness of BoN-aware fine-tuning in terms of improved performance and inference-time compute. In particular, we show that our methods improve the BoN performance of Gemma 2B on Hendrycks MATH from 26.8% to 30.8%, and Pass@K from 60% to 67%.

Self-correction is a highly desirable capability of large language models (LLMs), yet it has consistently been found to be largely ineffective in modern LLMs. Current methods for training self-correction typically depend on either multiple models, a more advanced model, or additional forms of supervision. To address these shortcomings, we develop a multi-turn online reinforcement learning (RL) approach, SCoRe, that significantly improves an LLM's self-correction ability using entirely self-generated data. To build SCoRe, we first show that variants of supervised fine-tuning (SFT) on offline model-generated correction traces are insufficient for instilling self-correction behavior. In particular, we observe that training via SFT either suffers from a distribution mismatch between the training data and the model's own responses or implicitly prefers only a certain mode of correction behavior that is often not effective at test time. SCoRe addresses these challenges by training under the model's own distribution of self-generated correction traces and using appropriate regularization to steer the learning process into learning a self-correction strategy that is effective at test time as opposed to simply fitting high-reward responses for a given prompt. This regularization prescribes running a first phase of RL on a base model to generate a policy initialization that is less susceptible to collapse and then using a reward bonus to amplify self-correction during training. When applied to Gemini 1.0 Pro and 1.5 Flash models, we find that SCoRe achieves state-of-the-art self-correction performance, improving the base models' self-correction by 15.6% and 9.1% respectively on the MATH and HumanEval benchmarks.

Efficient video tokenization remains a key bottleneck in learning general purpose vision models that are capable of processing long video sequences. Prevailing approaches are restricted to encoding videos to a fixed number of tokens, where too few tokens will result in overly lossy encodings, and too many tokens will result in prohibitively long sequence lengths. In this work, we introduce ElasticTok, a method that conditions on prior frames to adaptively encode a frame into a variable number of tokens. To enable this in a computationally scalable way, we propose a masking technique that drops a random number of tokens at the end of each frames's token encoding. During inference, ElasticTok can dynamically allocate tokens when needed -- more complex data can leverage more tokens, while simpler data only needs a few tokens. Our empirical evaluations on images and video demonstrate the effectiveness of our approach in efficient token usage, paving the way for future development of more powerful multimodal models, world models, and agents. Video examples of using ElasticTok can be found on our website: http://largeworldmodel.github.io/elastictok

The progress of autonomous web navigation has been hindered by the dependence on billions of exploratory interactions via online reinforcement learning, and domain-specific model designs that make it difficult to leverage generalization from rich out-of-domain data. In this work, we study data-driven offline training for web agents with vision-language foundation models. We propose an instruction-following multimodal agent, WebGUM, that observes both webpage screenshots and HTML pages and outputs web navigation actions, such as click and type. WebGUM is trained by jointly finetuning an instruction-finetuned language model and a vision encoder with temporal and local perception on a large corpus of demonstrations. We empirically demonstrate this recipe improves the agent's ability of grounded multimodal perception, HTML comprehension, and multi-step reasoning, outperforming prior works by a significant margin. On the MiniWoB, we improve over the previous best offline methods by more than 45.8%, even outperforming online-finetuned SoTA, humans, and GPT-4-based agent. On the WebShop benchmark, our 3-billion-parameter model achieves superior performance to the existing SoTA, PaLM-540B. Furthermore, WebGUM exhibits strong positive transfer to the real-world planning tasks on the Mind2Web. We also collect 347K high-quality demonstrations using our trained models, 38 times larger than prior work, and make them available to promote future research in this direction.

Pre-trained large language models (LLMs) have recently achieved better generalization and sample efficiency in autonomous web automation. However, the performance on real-world websites has still suffered from (1) open domainness, (2) limited context length, and (3) lack of inductive bias on HTML. We introduce WebAgent, an LLM-driven agent that learns from self-experience to complete tasks on real websites following natural language instructions. WebAgent plans ahead by decomposing instructions into canonical sub-instructions, summarizes long HTML documents into task-relevant snippets, and acts on websites via Python programs generated from those. We design WebAgent with Flan-U-PaLM, for grounded code generation, and HTML-T5, new pre-trained LLMs for long HTML documents using local and global attention mechanisms and a mixture of long-span denoising objectives, for planning and summarization. We empirically demonstrate that our modular recipe improves the success on real websites by over 50%, and that HTML-T5 is the best model to solve various HTML understanding tasks; achieving 18.7% higher success rate than the prior method on MiniWoB web automation benchmark, and SoTA performance on Mind2Web, an offline task planning evaluation.

Value functions are an essential component in deep reinforcement learning (RL), that are typically trained via mean squared error regression to match bootstrapped target values. However, scaling value-based RL methods to large networks has proven challenging. This difficulty is in stark contrast to supervised learning: by leveraging a cross-entropy classification loss, supervised methods have scaled reliably to massive networks. Observing this discrepancy, in this paper, we investigate whether the scalability of deep RL can also be improved simply by using classification in place of regression for training value functions. We show that training value functions with categorical cross-entropy significantly enhances performance and scalability across various domains, including single-task RL on Atari 2600 games, multi-task RL on Atari with large-scale ResNets, robotic manipulation with Q-transformers, playing Chess without search, and a language-agent Wordle task with high-capacity Transformers, achieving state-of-the-art results on these domains. Through careful analysis, we show that categorical cross-entropy mitigates issues inherent to value-based RL, such as noisy targets and non-stationarity. We argue that shifting to categorical cross-entropy for training value functions can substantially improve the scalability of deep RL at little-to-no cost.

We propose a framework for classifying the capabilities and behavior of Artificial General Intelligence (AGI) models and their precursors. This framework introduces levels of AGI performance, generality, and autonomy, providing a common language to compare models, assess risks, and measure progress along the path to AGI. To develop our framework, we analyze existing definitions of AGI, and distill six principles that a useful ontology for AGI should satisfy. With these principles in mind, we propose “Levels of AGI” based on depth (performance) and breadth (generality) of capabilities, and reflect on how current systems fit into this ontology. We discuss the challenging requirements for future benchmarks that quantify the behavior and capabilities of AGI models against these levels. Finally, we discuss how these levels of AGI interact with deployment considerations such as autonomy and risk, and emphasize the importance of carefully selecting Human-AI Interaction paradigms for responsible and safe deployment of highly capable AI systems.

Large language models (LLMs) excel at few-shot in-context learning (ICL) -- learning from a few examples provided in context at inference, without any weight updates. Newly expanded context windows allow us to investigate ICL with hundreds or thousands of examples – the many-shot regime. Going from few-shot to many-shot, we observe significant performance gains across a wide variety of generative and discriminative tasks. While promising, many-shot ICL can be bottlenecked by the available amount of human-generated outputs. To mitigate this limitation, we explore two new settings: (1) "Reinforced ICL" that uses model-generated chain-of-thought rationales in place of human rationales, and (2) "Unsupervised ICL" where we remove rationales from the prompt altogether, and prompts the model only with domain-specific inputs. We find that both Reinforced and Unsupervised ICL can be quite effective in the many-shot regime, particularly on complex reasoning tasks. We demonstrate that, unlike few-shot learning, many-shot learning is effective at overriding pretraining biases, can learn high-dimensional functions with numerical inputs, and performs comparably to supervised fine-tuning. Finally, we reveal the limitations of next-token prediction loss as an indicator of downstream ICL performance.

Many algorithms for aligning LLMs with human preferences assume that human preferences are binary and deterministic. However, human preferences can vary across individuals, and therefore should be represented distributionally. In this work, we introduce the distributional soft preference labels and improve Direct Preference Optimization (DPO) with a weighted geometric average of the LLM output likelihood in the loss function. This approach adjusts the scale of learning loss based on the soft labels such that the loss would approach zero when the responses are closer to equally preferred. This simple modification can be easily applied to any DPO-based methods and mitigate over-optimization and objective mismatch, which prior works suffer from. Our experiments simulate the soft preference labels with AI feedback from LLMs and demonstrate that geometric averaging consistently improves performance on standard benchmarks for alignment research. In particular, we observe more preferable responses than binary labels and significant improvements where modestly-confident labels are in the majority.

Reinforcement Learning (RL) algorithms are often known for sample inefficiency and difficult generalization. Recently, Unsupervised Environment Design (UED) emerged as a new paradigm for zero-shot generalization by simultaneously learning a task distribution and agent policies on the generated tasks. This is a non-stationary process where the task distribution evolves along with agent policies; creating an instability over time. While past works demonstrated the potential of such approaches, sampling effectively from the task space remains an open challenge, bottlenecking these approaches. To this end, we introduce CLUTR: a novel unsupervised curriculum learning algorithm that decouples task representation and curriculum learning into a two-stage optimization. It first trains a recurrent variational autoencoder on randomly generated tasks to learn a latent task manifold. Next, a teacher agent creates a curriculum by maximizing a minimax REGRET-based objective on a set of latent tasks sampled from this manifold. Using the fixed-pretrained task manifold, we show that CLUTR successfully overcomes the non-stationarity problem and improves stability. Our experimental results show CLUTR outperforms PAIRED, a principled and popular UED method, in the challenging CarRacing and navigation environments: achieving 10.6X and 45% improvement in zero-shot generalization, respectively. CLUTR also performs comparably to the non-UED state-of-the-art for CarRacing, while requiring 500X fewer environment interactions. We open source our code at https://github.com/clutr/clutr.

Motion planning tasks are often innately invariant to certain geometric transformations, or in other words, symmetric. This property, however, is not always reflected in learned policies that are trained on these tasks. Although this asymmetry can be addressed through data augmentation or additional training samples, doing so comes at a cost of increased training time. Instead of trying to remedy this issue during the learning process, we leverage this disparity during execution. We propose the symmetry exploitation policy, an augmentation in the post-hoc execution stage of RL policies. During the planning stage, we present the learned policy with an invariant, geometrically transformed version of the observation as an alternate perspective of the state. This allows the policy to produce multiple possible actions for a single state, and choose the action with the highest estimated value. Unlike other symmetry exploitation methods for learning solutions in motion planning, this method completely bypasses the need for additional training. We show the effect of the symmetry exploitation policy on DQN, A2C, and PPO policies, in three motion problems with different dimensions, observation types, and symmetries. The results show that by exploiting the symmetry of the task, a trained model achieves improved performance and better generalization, and can achieve comparable results to retraining, augmentation, or extended training, without incurring any additional training time. The efficacy is most prominent in more complex tasks, as 89 of the 100 models involved in the case study improve when using the method.

Imitation learning (IL) is a simple and powerful way to use high-quality human driving data, which can be collected at scale, to produce human-like behavior. However, policies based on imitation learning alone often fail to sufficiently account for safety and reliability concerns. In this paper, we show how imitation learning combined with reinforcement learning using simple rewards can substan-tially improve the safety and reliability of driving policies over those learned from imitation alone. In particular, we train a policy on over lOOk miles of urban driving data, and measure its effectiveness in test scenarios grouped by different levels of collision likelihood. Our analysis shows that while imitation can perform well in low-difficulty scenarios that are well-covered by the demonstration data, our proposed approach significantly improves robustness on the most challenging scenarios (over 38 % reduction in failures). To our knowledge, this is the first application of a combined imitation and reinforcement learning approach in autonomous driving that utilizes large amounts of real- world human driving data.

Machine learning (ML) has become a prevalent approach to tame the complexity of design space exploration for domain-specific architectures. While appealing, using ML for design space exploration poses several challenges. First, it is not straightforward to identify the most suitable algorithm from an ever-increasing pool of ML methods. Second, assessing the trade-offs between performance and sample efficiency across these methods is inconclusive. Finally, the lack of a holistic framework for fair, reproducible, and objective comparison across these methods hinders the progress of adopting ML-aided architecture design space exploration and impedes creating repeatable artifacts. To mitigate these challenges, we introduce ArchGym, an open-source gymnasium and easy-to-extend framework that connects a diverse range of search algorithms to architecture simulators. To demonstrate its utility, we evaluate ArchGym across multiple vanilla and domain-specific search algorithms in the design of a custom memory controller, deep neural network accelerators, and a custom SoC for AR/VR workloads, collectively encompassing over 21K experiments. The results suggest that with an unlimited number of samples, ML algorithms are equally favorable to meet the user-defined target specification if its hyperparameters are tuned thoroughly; no one solution is necessarily better than another (e.g., reinforcement learning vs. Bayesian methods). We coin the term "hyperparameter lottery" to describe the relatively probable chance for a search algorithm to find an optimal design provided meticulously selected hyperparameters. Additionally, the ease of data collection and aggregation in ArchGym facilitates research in ML-aided architecture design space exploration. As a case study, we show this advantage by developing a proxy cost model with an RMSE of 0.61% that offers a 2,000-fold reduction in simulation time. Code and data for ArchGym is available at https://bit.ly/ArchGym.

Simulation is an essential tool to develop and benchmark autonomous vehicle planning software in a safe and cost-effective manner. However, realistic simulation requires accurate modeling of multi-agent interactive behaviors to be trustworthy, behaviors which can be highly nuanced and complex. To address these challenges, we introduce Waymax, a new data-driven simulator for autonomous driving in multi-agent scenes, designed for large-scale simulation and testing. Waymax uses publicly-released, real-world driving data (e.g., the Waymo Open Motion Dataset) to initialize or play back a diverse set of multi-agent simulated scenarios. It runs entirely on hardware accelerators such as TPUs/GPUs and supports in-graph simulation for training, making it suitable for modern large-scale, distributed machine learning workflows. To support online training and evaluation, Waymax includes several learned and hard-coded behavior models that allow for realistic interaction within simulation. To supplement Waymax, we benchmark a suite of popular imitation and reinforcement learning algorithms with ablation studies on different design decisions, where we highlight the effectiveness of routes as guidance for planning agents and the ability of RL to overfit against simulated agents.

We study the automatic generation of navigation instructions from 360° images captured on indoor routes. Existing generators suffer from poor visual grounding, causing them to rely on language priors and hallucinate objects. Our Marky-mt5 system addresses this by focusing on visual landmarks; it comprises a first stage landmark detector and a second stage generator-a multimodal, multilingual, multi-task encoder-decoder. To train it, we bootstrap grounded landmark annotations on top of the Room-across-Room (RxR) dataset. Using text parsers, weak supervision from RxR's pose traces, and a multilingual image-text encoder trained on 1.8b images, we identify 971k English, Hindi and Telugu landmark descriptions and ground them to specific regions in panoramas. On Room-to-Room, human wayfind-ers obtain success rates (SR) of 71% following Marky-mt5's instructions, just shy of their 75% SR following human instructions-and well above SRs with other genera-tors. Evaluations on RxR's longer, diverse paths obtain 61-64% SRs on three languages. Generating such high-quality navigation instructions in novel environments is a step to-wards conversational navigation tools and could facilitate larger-scale training of instruction-following agents.

Enabling robots to solve multiple manipulation tasks has a wide range of industrial applications. While learning-based approaches enjoy flexibility and generalizability, scaling these approaches to solve such compositional tasks remains a challenge. In this work, we aim to solve multi-task learning through the lens of sequence-conditioning and weighted sampling. First, we propose a new suite of benchmark specifically aimed at compositional tasks, MultiRavens, which allows defining custom task combinations through task modules that are inspired by industrial tasks and exemplify the difficulties in vision-based learning and planning methods. Second, we propose a vision-based end-to-end system architecture, Sequence-Conditioned Transporter Networks, which augments Goal-Conditioned Transporter Networks with sequence-conditioning and weighted sampling and can efficiently learn to solve multi-task long horizon problems. Our analysis suggests that not only the new framework significantly improves pick-and-place performance on novel 10 multi-task benchmark problems, but also the multi-task learning with weighted sampling can vastly improve learning and agent performances on individual tasks.

We introduce an early-phase bottleneck analysis and characterization model called the F-1 for designing computing systems that target autonomous Unmanned Aerial Vehicles (UAVs). The model provides insights by exploiting the fundamental relationships between various components in the autonomous UAV, such as sensor, compute, and body dynamics. To guarantee safe operation while maximizing the performance (e.g., velocity) of the UAV, the compute, sensor, and other mechanical properties must be carefully selected or designed. The F-1 model provides visual insights that can aid a system architect in understanding the optimal compute design or selection for autonomous UAVs. The model is experimentally validated using real UAVs, and the error is between 5.1% to 9.5% compared to real-world flight tests. An interactive web-based tool for the F-1 model called Skyline is available for free of cost use at: https://bit.ly/skyline-tool

Building domain-specific accelerators is becoming increasingly paramount to meet the high-performance requirements under stringent power and real-time constraints. However, emerging application domains like autonomous vehicles are complex systems with constraints extending beyond the computing stack. Manually selecting and navigating the design space to design custom and efficient domain-specific SoCs (DSSoC) is tedious and expensive. Hence, there is a need for automated DSSoC design methodologies. In this paper, we use agile and autonomous UAVs as a case study to understand how to automate domain-specific SoCs design for autonomous vehicles. Architecting a UAV DSSoC requires consideration of parameters such as sensor rate, compute throughput, and other physical characteristics (e.g., payload weight, thrust-to-weight ratio) that affect overall performance. Iterating over several component choices results in a combinatorial explosion of the number of possible combinations: from tens of thousands to billions, depending on implementation details. To navigate the DSSoC design space efficiently, we introduce AutoPilot, a systematic methodology for automatically designing DSSoC for autonomous UAVs. AutoPilot uses machine learning to navigate the large DSSoC design space and automatically select a combination of autonomy algorithm and hardware accelerator while considering the cross-product effect across different UAV components. AutoPilot consistently outperforms general-purpose hardware selections like Xavier NX and Jetson TX2, as well as dedicated hardware accelerators built for autonomous UAVs. DSSoC designs generated by AutoPilot increase the number of missions on average by up to 2.25×, 1.62×, and 1.43× for nano, micro, and mini-UAVs, respectively, over baselines. Further, we discuss the potential application of AutoPilot methodology to other related autonomous vehicles.

We tackle real-world problems with complex structures beyond the pixel-based game or simulator. We formulate it as a few-shot reinforcement learning problem where a task is characterized by a subtask graph that defines a set of subtasks and their dependencies that are unknown to the agent. Different from the previous meta-RL methods trying to directly infer the unstructured task embedding, our multi-task subtask graph inferencer (MTSGI) first infers the common high-level task structure in terms of the subtask graph from the training tasks, and use it as a prior to improve the task inference in testing. Our experiment results on 2D grid-world and complex web navigation domains show that the proposed method can learn and leverage the common underlying structure of the tasks for faster adaptation to the unseen tasks than various existing algorithms such as meta reinforcement learning, hierarchical reinforcement learning, and other heuristic agents.

The combination of Reinforcement Learning (RL) with deep learning has led to a series of impressive feats, with many believing (deep) RL provides a path towards generally capable agents. However, the success of RL agents is often highly sensitive to design choices in the training process, which may require tedious and error-prone manual tuning. This makes it challenging to use RL for new problems and also limits its full potential. In many other areas of machine learning, AutoML has shown that it is possible to automate such design choices, and AutoML has also yielded promising initial results when applied to RL. However, Automated Reinforcement Learning (AutoRL) involves not only standard applications of AutoML but also includes additional challenges unique to RL, that naturally produce a different set of methods. As such, AutoRL has been emerging as an important area of research in RL, providing promise in a variety of applications from RNA design to playing games, such as Go. Given the diversity of methods and environments considered in RL, much of the research has been conducted in distinct subfields, ranging from meta-learning to evolution. In this survey, we seek to unify the field of AutoRL, provide a common taxonomy, discuss each area in detail and pose open problems of interest to researchers going forward.

Motion planning among dynamic obstacles is an essential capability towards navigation in the real-world. Sampling-based motion planning algorithms find solutions by approximating the robot’s configuration space through a graph representation, predicting or computing obstacles’ trajectories, and finding feasible paths via a pathfinding algorithm. In this work, we seek to improve the performance of these subproblems by identifying regions critical to dynamic environment navigation and leveraging them to construct sparse probabilistic roadmaps. Motion planning and pathfinding algorithms should allow robots to prevent encounters with obstacles, irrespective of their trajectories, by being conscious of spatial context cues such as the location of chokepoints (e.g., doorways). Thus, we propose a self-supervised methodology for learning to identify regions frequently used for obstacle avoidance from local environment features. As an application of this concept, we leverage a neural network to generate hierarchical probabilistic roadmaps termed Avoidance Critical Probabilistic Roadmaps (ACPRM). These roadmaps contain motion structures that enable efficient obstacle avoidance, reduce the search and planning space, and increase a roadmap’s reusability and coverage. ACPRMs are demonstrated to achieve up to five orders of magnitude improvement over grid-sampling in the multi-agent setting and up to ten orders of magnitude over a competitive baseline in the multi-query setting.

We present fully autonomous source seeking onboard a highly constrained nano quadcopter, by contributing application-specific system and observation feature design to enable inference of a deep-RL policy onboard a nano quadcopter. Our deep-RL algorithm finds a high-performance solution to a challenging problem, even in presence of high noise levels and generalizes across real and simulation environments with different obstacle configurations. We verify our approach with simulation and in-field testing on a Bitcraze CrazyFlie using only the cheap and ubiquitous Cortex-M4 microcontroller unit. The results show that by end-to-end application-specific system design, our contribution consumes almost three times less additional power, as compared to a competitive learning-based navigation approach onboard a nano quadcopter. Thanks to our observation space, which we carefully design within the resource constraints, our solution achieves a 94% success rate in cluttered and randomized test environments, as compared to the previously achieved 80%. We also compare our strategy to a simple finite state machine (FSM), geared towards efficient exploration, and demonstrate that our policy is more robust and resilient at obstacle avoidance as well as up to 70% more efficient in source seeking. To this end, we contribute a cheap and lightweight end- to-end tiny robot learning (tinyRL) solution, running onboard a nano quadcopter, that proves to be robust and efficient in a challenging task.

Many real-world problems are compositional – solving them requires completing interdependent sub-tasks, either in series or in parallel, that can be represented as a dependency graph. Deep reinforcement learning (RL) agents often struggle to learn such complex tasks due to the long time horizons and sparse rewards. To address this problem, we present Compositional Design of Environments (CoDE), which trains a Generator agent to automatically build a series of compositional tasks tailored to the RL agent’s current skill level. This automatic curriculum not only enables the agent to learn more complex tasks than it could have otherwise, but also selects tasks where the agent’s performance is weak, enhancing its robustness and ability to generalize zero-shot to unseen tasks at test-time. We analyze why current environment generation techniques are insufficient for the problem of generating compositional tasks, and propose a new algorithm that addresses these issues. Our results assess learning and generalization across multiple compositional tasks, including the real-world problem of learning to navigate and interact with web pages. We learn to generate environments composed of multiple pages or rooms, and train RL agents capable of completing wide-range of complex tasks in those environments. We contribute two new benchmark frameworks for generating compositional tasks, compositional MiniGrid and gMiniWoB for web navigation. CoDE yields 4x higher success rate than the strongest baseline, and demonstrates strong performance of real websites learned on 3500 primitive tasks.

Abstract We introduce Air Learning, an open-source simulator, and a gym environment for deep reinforcement learning research on resource-constrained aerial robots. Equipped with domain randomization, Air Learning exposes a UAV agent to a diverse set of challenging scenarios. We seed the toolset with point-to-point obstacle avoidance tasks in three different environments and Deep Q Networks (DQN) and Proximal Policy Optimization (PPO) trainers. Air Learning assesses the policies’ performance under various quality-of-flight (QoF) metrics, such as the energy consumed, endurance, and the average trajectory length, on resource-constrained embedded platforms like a Raspberry Pi. We find that the trajectories on an embedded Ras-Pi are vastly different from those predicted on a high-end desktop system, resulting in up to $$40\%$$ 40% longer trajectories in one of the environments. To understand the source of such discrepancies, we use Air Learning to artificially degrade high-end desktop performance to mimic what happens on a low-end embedded system. We then propose a mitigation technique that uses the hardware-in-the-loop to determine the latency distribution of running the policy on the target platform (onboard compute on aerial robot). A randomly sampled latency from the latency distribution is then added as an artificial delay within the training loop. Training the policy with artificial delays allows us to minimize the hardware gap (discrepancy in the flight time metric reduced from 37.73% to 0.5%). Thus, Air Learning with hardware-in-the-loop characterizes those differences and exposes how the onboard compute’s choice affects the aerial robot’s performance. We also conduct reliability studies to assess the effect of sensor failures on the learned policies. All put together, Air Learning enables a broad class of deep RL research on UAVs. The source code is available at: https://github.com/harvard-edge/AirLearning .

Autonomous and mobile cyber-physical machines are becoming an inevitable part of our future. In particular, Micro Aerial Vehicles (MAVs) have seen a resurgence in activity. With multiple use cases, such as surveillance, search and rescue, package delivery, and more, these unmanned aerial systems are on the cusp of demonstrating their full potential. Despite such promises, these systems face many challenges, one of the most prominent of which is their low endurance caused by their limited onboard energy. Since the success of a mission depends on whether the drone can finish it within such duration and before it runs out of battery, improving both the time and energy associated with the mission are of high importance. Such improvements have traditionally been arrived at through the use of better algorithms. But our premise is that more powerful and efficient onboard compute can also address the problem. In this article, we investigate how the compute subsystem, in a cyber-physical mobile machine such as a Micro Aerial Vehicle, can impact mission time (time to complete a mission) and energy. Specifically, we pose the question as what is the role of computing for cyber-physical mobile robots? We show that compute and motion are tightly intertwined, and as such a close examination of cyber and physical processes and their impact on one another is necessary. We show different “impact paths” through which compute impacts mission metrics and examine them using a combination of analytical models, simulation, and micro and end-to-end benchmarking. To enable similar studies, we open sourced MAVBench , our tool-set, which consists of (1) a closed-loop real-time feedback simulator and (2) an end-to-end benchmark suite composed of state-of-the-art kernels. By combining MAVBench, analytical modeling, and an understanding of various compute impacts, we show up to 2X and 1.8X improvements for mission time and mission energy for two optimization case studies, respectively. Our investigations, as well as our optimizations, show that cyber-physical co-design, a methodology with which both the cyber and physical processes/quantities of the robot are developed with consideration of one another, similar to hardware-software co-design, is necessary for arriving at the design of the optimal robot.

Sampling-based motion planning techniques have emerged as an efficient algorithmic paradigm for solving complex motion planning problems. These approaches use a set of probing samples to construct an implicit graph representation of the robot's state space, allowing arbitrarily accurate representations as the number of samples increases to infinity. In practice, however, solution trajectories only rely on a few critical states, often defined by structure in the state space (e.g., doorways). In this work we propose a general method to identify these critical states via graph-theoretic techniques (betweenness centrality) and learn to predict criticality from only local environment features. These states are then leveraged more heavily via global connections within a hierarchical graph, termed Critical Probabilistic Roadmaps. Critical PRMs are demonstrated to achieve up to three orders of magnitude improvement over uniform sampling, while preserving the guarantees and complexity of sampling-based motion planning. A video is available at https://youtu.be/AYoD-pGd9ms.

Imitation learning is a popular approach for training effective visual navigation policies. However, collecting expert demonstrations for legged robots is challenging as these robots can be hard to control, move slowly, and cannot operate continuously for long periods of time. In this work, we propose a zero-shot imitation learning framework for training a goal-driven visual navigation policy on a legged robot from human demonstrations (third-person perspective), allowing for high-quality navigation and cost-effective data collection. However, imitation learning from third-person demonstrations raises unique challenges. First, these demonstrations are captured from different camera perspectives, which we address via a feature disentanglement network (FDN) that extracts perspective-invariant state features. Second, as transition dynamics vary between systems, we reconstruct missing action labels by either building an inverse model of the robot's dynamics in the feature space and applying it to the human demonstrations or developing a Graphic User Interface (GUI) to label human demonstrations. To train a navigation policy we use a model-based imitation learning approach with FDN and action-labeled human demonstrations. We show that our framework can learn an effective policy for a legged robot, Laikago, from human demonstrations in both simulated and real-world environments. Our approach is zero-shot as the robot never navigates the same paths during training as those at testing time. We justify our framework by performing a comparative study.

Long-range indoor navigation requires guiding robots with noisy sensors and controls through cluttered environments along paths that span a variety of buildings. We achieve this with PRM-RL, a hierarchical robot navigation method in which reinforcement learning (RL) agents that map noisy sensors to robot controls learn to solve short-range obstacle avoidance tasks, and then sampling-based planners map where these agents can reliably navigate in simulation; these roadmaps and agents are then deployed on robots, guiding them along the shortest path where the agents are likely to succeed. In this article, we use probabilistic roadmaps (PRMs) as the sampling-based planner, and AutoRL as the RL method in the indoor navigation context. We evaluate the method with a simulation for kinematic differential drive and kinodynamic car-like robots in several environments, and on differential-drive robots at three physical sites. Our results show that PRM-RL with AutoRL is more successful than several baselines, is robust to noise, and can guide robots over hundreds of meters in the face of noise and obstacles in both simulation and on robots, including over 5.8 km of physical robot navigation.

Deep Reinforcement Learning (RL) has recently emerged as a solution for moving obstacle avoidance. Deep RL learns to simultaneously predict obstacle motions and corresponding avoidance actions directly from robot sensors, even for obstacles with different dynamics models. However, deep RL methods typically cannot guarantee policy convergences, i.e., cannot provide probabilistic collision avoidance guarantees. In contrast, stochastic reachability (SR), a computationally expensive formal method that employs a known obstacle dynamics model, identifies the optimal avoidance policy and provides strict convergence guarantees. The availability of the optimal solution for versions of the moving obstacle problem provides a baseline to compare trained deep RL policies. In this paper, we compare the expected cumulative reward and actions of these policies to SR, and find the following. 1) The state-value function approximates the optimal collision probability well, thus explaining the high empirical performance. 2) RL policies deviate from the optimal significantly thus negatively impacting collision avoidance in some cases. 3) Evidence suggests that the deviation is caused, at least partially, by the actor net failing to approximate the action corresponding to the highest state-action value.

We present PRM-RL, a hierarchical method for long-range navigation task completion that combines sampling-based path planning with reinforcement learning (RL). The RL agents learn short-range, point-to-point navigation policies that capture robot dynamics and task constraints without knowledge of the large-scale topology. Next, the sampling-based planners provide roadmaps which connect robot configurations that can be successfully navigated by the RL agent. The same RL agents are used to control the robot under the direction of the planning, enabling long-range navigation. We use the Probabilistic Roadmaps (PRMs) for the sampling-based planner. The RL agents are constructed using feature-based and deep neural net policies in continuous state and action spaces. We evaluate PRM-RL, both in simulation and on-robot, on two navigation tasks with non-trivial robot dynamics: end-to-end differential drive indoor navigation in office environments, and aerial cargo delivery in urban environments with load displacement constraints. Our results show improvement in task completion over both RL agents on their own and traditional sampling-based planners. In the indoor navigation task, PRM-RL successfully completes up to 215 m long trajectories under noisy sensor conditions, and the aerial cargo delivery completes flights over 1000 m without violating the task constraints in an environment 63 million times larger than used in training.

Unmanned Aerial Vehicles (UAVs) are getting closer to becoming ubiquitous in everyday life. Among them, Micro Aerial Vehicles (MAVs) have seen an outburst of attention recently, specifically in the area with a demand for autonomy. A key challenge standing in the way of making MAVs autonomous is that researchers lack the comprehensive understanding of how performance, power, and computational bottlenecks affect MAV applications. MAVs must operate under a stringent power budget, which severely limits their flight endurance time. As such, there is a need for new tools, benchmarks, and methodologies to foster the systematic development of autonomous MAVs. In this paper, we introduce the MAVBench' framework which consists of a closed-loop simulator and an end-to-end application benchmark suite. A closed-loop simulation platform is needed to probe and understand the intra-system (application data flow) and inter-system (system and environment) interactions in MAV applications to pinpoint bottlenecks and identify opportunities for hardware and software co-design and optimization. In addition to the simulator, MAVBench provides a benchmark suite, the first of its kind, consisting of a variety of MAV applications designed to enable computer architects to perform characterization and develop future aerial computing systems. Using our open source, end-to-end experimental platform, we uncover a hidden, and thus far unexpected compute to total system energy relationship in MAVs. Furthermore, we explore the role of compute by presenting three case studies targeting performance, energy and reliability. These studies confirm that an efficient system design can improve MAV's battery consumption by up to 1.8X.

Manual derivation of optimal robot motions for task completion is difficult, especially when a robot is required to balance its actions between opposing preferences. One solution has been proposed to automatically learn near optimal motions with Reinforcement Learning (RL). This has been successful for several tasks including swing-free UAV flight, table tennis, and autonomous driving. However, high-dimensional problems remain a challenge. We address this dimensionality constraint with PrEference Appraisal Reinforcement Learning (PEARL), which solves tasks with opposing preferences for acceleration controlled robots. PEARL projects the high-dimensional continuous robot state space to a low dimensional preference feature space resulting in efficient and adaptable planning. We demonstrate that on a dynamic obstacle avoidance robotic task, a single learning on a much simpler problem performs real-time decision-making for significantly larger, high-dimensional problems working in unbounded continuous states and actions. We trained the agent with 4 static obstacles, while the trained agent avoids up to 900 moving obstacles with complex hybrid stochastic obstacle dynamics in a highly constrained space using only limited information about the environment. We compare these tasks to traditional, often manually tuned solutions for these high-dimensional problems.

Physical stochastic disturbances, such as wind, often affect the motion of robots that perform complex tasks in real-world conditions. These disturbances pose a control challenge because resulting drift induces uncertainty and changes in the robot's speed and direction. This paper presents an online control policy based on supervised machine learning, Least Squares Axial Sum Policy Approximation (LSAPA), that generates trajectories for robotic preference-balancing tasks under stochastic disturbances. The task is learned offline with reinforcement learning, assuming no disturbances, and then trajectories are planned online in the presence of disturbances using the current observed information. We model the robot as a stochastic control-affine system with unknown dynamics impacted by a Gaussian process, and the task as a continuous Markov Decision Process. Replacing a traditional greedy policy, LSAPA works for high-dimensional control-affine systems impacted by stochastic disturbances and is linear in the input dimensionality. We verify the method for Swing-free Aerial Cargo Delivery and Rendezvous tasks. Results show that LSAPA selects an input an order of magnitude faster than comparative methods, rejecting a range of stochastic disturbances. Further, experiments on a quadrotor demonstrate that LSAPA trajectories that are suitable for physical systems.

Attaining autonomous flight is an important task in aerial robotics. Often flight trajectories are not only subject to unknown system dynamics, but also to specific task constraints. This paper presents a motion planning method for generating trajectories with minimal residual oscillations (swing-free) for rotorcraft carrying a suspended loads. We rely on a finite-sampling, batch reinforcement learning algorithm to train the system for a particular load. We find criteria that allow the trained agent to be transferred to a variety of models, state and action spaces and produce a number of different trajectories. Through a combination of simulations and experiments, we demonstrate that the inferred policy is robust to noise and the unmodeled dynamics of the system. The contributions of this work are 1) applying reinforcement learning to solve the problem of finding swing-free trajectories for rotorcraft, 2) designing a problem-specific feature vector for value function approximation, 3) giving sufficient conditions for successful learning transfer to different models, state and action spaces, and 4) verification of the resulting trajectories in both simulation and autonomous control of quadrotors with suspended loads.

In this paper, we present a problem where a suspended load, carried by a rotorcraft aerial robot, performs trajectory tracking. We want to accomplish this by specifying the reference trajectory for the suspended load only. The aerial robot needs to discover/learn its own trajectory which ensures that the suspended load tracks the reference trajectory. As a solution, we propose a method based on least-square policy iteration (LSPI) which is a type of reinforcement learning algorithm. The proposed method is verified through simulation and experiments.