Spatial intelligence (SI) represents a cognitive ability encompassing the visualization, manipulation, and reasoning about spatial relationships, underpinning disciplines from neuroscience to robotics. We introduce SITE, a benchmark dataset towards SI Thorough Evaluation in a standardized format of multi-choice visual question-answering, designed to assess large vision-language models' spatial intelligence across diverse visual modalities (single-image, multi-image, and video) and SI factors (figural to environmental scales, spatial visualization and orientation, intrinsic and extrinsic, static and dynamic). Our approach to curating the benchmark combines a bottom-up survey about 31 existing datasets and a top-down strategy drawing upon three classification systems in cognitive science, which prompt us to design two novel types of tasks about view-taking and dynamic scenes. Extensive experiments reveal that leading models fall behind human experts especially in spatial orientation, a fundamental SI factor. Moreover, we demonstrate a positive correlation between a model's spatial reasoning proficiency and its performance on an embodied AI task.

Robot learning requires a considerable amount of high-quality data to realize the promise of generalization. However, large data sets are costly to collect in the real world. Physics simulators can cheaply generate vast data sets with broad coverage over states, actions, and environments. However, physics engines are fundamentally misspecified approximations to reality. This makes direct zero-shot transfer from simulation to reality challenging, especially in tasks where precise and force-sensitive manipulation is necessary. Thus, fine-tuning these policies with small real-world data sets is an appealing pathway for scaling robot learning. However, current reinforcement learning fine-tuning frameworks leverage general, unstructured exploration strategies which are too inefficient to make real-world adaptation practical. This paper introduces the \emph{Simulation-Guided Fine-tuning} (SGFT) framework, which demonstrates how to extract structural priors from physics simulators to substantially accelerate real-world adaptation. Specifically, our approach uses a value function learned in simulation to guide real-world exploration. We demonstrate this approach across five real-world dexterous manipulation tasks where zero-shot sim-to-real transfer fails. We further demonstrate our framework substantially outperforms baseline fine-tuning methods, requiring up to an order of magnitude fewer real-world samples and succeeding at difficult tasks where prior approaches fail entirely. Last but not least, we provide theoretical justification for this new paradigm which underpins how SGFT can rapidly learn high-performance policies in the face of large sim-to-real dynamics gaps.

Although large vision-language-action (VLA) models pretrained on extensive robot datasets offer promising generalist policies for robotic learning, they still struggle with spatial-temporal dynamics in interactive robotics, making them less effective in handling complex tasks, such as manipulation. In this work, we introduce visual trace prompting, a simple yet effective approach to facilitate VLA models’ spatial-temporal awareness for action prediction by encoding state-action trajectories visually. We develop a new TraceVLA model by finetuning OpenVLA on our own collected dataset of 150K robot manipulation trajectories using visual trace prompting. Evaluations of TraceVLA across 137 configurations in SimplerEnv and 4 tasks on a physical WidowX robot demonstrate state-of-the-art performance, outperforming OpenVLA by 10% on SimplerEnv and 3.5x on real-robot tasks and exhibiting robust generalization across diverse embodiments and scenarios. To further validate the effectiveness and generality of our method, we present a compact VLA model based on 4B Phi-3-Vision, pretrained on the Open-X-Embodiment and finetuned on our dataset, rivals the 7B OpenVLA baseline while significantly improving inference efficiency.

We propose **H**e**u**ristic **Bl**ending (HUBL), a simple performance-improving technique for a broad class of offline RL algorithms based on value bootstrapping. HUBL modifies the Bellman operators used in these algorithms, partially replacing the bootstrapped values with heuristic ones that are estimated with Monte-Carlo returns. For trajectories with higher returns, HUBL relies more on the heuristic values and less on bootstrapping; otherwise, it leans more heavily on bootstrapping. HUBL is very easy to combine with many existing offline RL implementations by relabeling the offline datasets with adjusted rewards and discount factors. We derive a theory that explains HUBL's effect on offline RL as reducing offline RL's complexity and thus increasing its finite-sample performance. Furthermore, we empirically demonstrate that HUBL consistently improves the policy quality of four state-of-the-art bootstrapping-based offline RL algorithms (ATAC, CQL, TD3+BC, and IQL), by 9% on average over 27 datasets of the D4RL and Meta-World benchmarks.

Temporal action abstractions, along with belief state representations, are a powerful knowledge sharing mechanism for sequential decision making. In this work, we propose a novel view that treats inducing temporal action abstractions as a sequence compression problem. To do so, we bring a subtle but critical component of LLM training pipelines -- input tokenization via byte pair encoding (BPE) -- to bear on the seemingly distant task of learning skills of variable time span in continuous control domains. We introduce an approach called Primitive Sequence Encoding (PRISE) that combines continuous action quantization with BPE to learn powerful action abstractions. We empirically show that high-level skills discovered by PRISE from a multitask set of robotic manipulation demonstrations significantly boost the learning performance of behavior cloning on downstream tasks.

Detecting air flows caused by phenomena such as heat convection is valuable in multiple scenarios, including leak identification and locating thermal updrafts for extending UAV flight duration. Unfortunately, the heat signature of these flows is often too subtle to be seen by a thermal camera. While convection also leads to fluctuations in air density and hence causes so-called schlieren – intensity and color variations in images – existing techniques such as Background-oriented schlieren (BOS) allow detecting them only against a known background and from a static camera, making these approaches unsuitable for moving vehicles. In this work we demonstrate the feasibility of visualizing air movement by predicting the corresponding schlieren-induced optical flow from a single greyscale image captured by a moving camera against an unfamiliar background. We first record and label a set of optical flows in an indoor setup using standard BOS techniques. We then train a convolutional neural network (CNN) by applying the previously collected optical flow distortions to a dataset containing a mixture of real and synthetically generated images to predict the two-dimensional optical flow from a single image. Finally, we evaluate our approach on the task of extracting the optical flow caused by schlieren from both a static and moving camera on previously unseen flow patterns and background images.

Only a small percentage of blind and low-vision people use traditional mobility aids such as a cane or a guide dog. Various assistive technologies have been proposed to address the limitations of traditional mobility aids. These devices often give either the user or the device majority of the control. In this work, we explore how varying levels of control affect the users' sense of agency, trust in the device, confidence, and successful navigation. We present Glide, a novel mobility aid with two modes for control: Glide-directed and User-directed. We employ Glide in a study (N=9) in which blind or low-vision participants used both modes to navigate through an indoor environment. Overall, participants found that Glide was easy to use and learn. Most participants trusted Glide despite its current limitations, and their confidence and performance increased as they continued to use Glide. Users' control mode preferences varied in different situations; no single mode "won" in all situations.

We present a novel observation about the behavior of offline reinforcement learning (RL) algorithms: on many benchmark datasets, offline RL can produce well-performing and safe policies even when trained with "wrong" reward labels, such as those that are zero everywhere or are negatives of the true rewards. This phenomenon cannot be easily explained by offline RL's return maximization objective. Moreover, it gives offline RL a degree of robustness that is uncharacteristic of its online RL counterparts, which are known to be sensitive to reward design. We demonstrate that this surprising robustness property is attributable to an interplay between the notion of *pessimism* in offline RL algorithms and certain implicit biases in common data collection practices. As we prove in this work, pessimism endows the agent with a *survival instinct*, i.e., an incentive to stay within the data support in the long term, while the limited and biased data coverage further constrains the set of survival policies. Formally, given a reward class -- which may not even contain the true reward -- we identify conditions on the training data distribution that enable offline RL to learn a near-optimal and safe policy from any reward within the class. We argue that the survival instinct should be taken into account when interpreting results from existing offline RL benchmarks and when creating future ones. Our empirical and theoretical results suggest a new paradigm for offline RL, whereby an agent is "nudged" to learn a desirable behavior with imperfect reward but purposely biased data coverage. Please visit our website [https://survival-instinct.github.io](https://survival-instinct.github.io) for accompanied code and videos.

Simulated humanoids are an appealing research domain due to their physical capabilities. Nonetheless, they are also challenging to control, as a policy must drive an unstable, discontinuous, and high-dimensional physical system. One widely studied approach is to utilize motion capture (MoCap) data to teach the humanoid agent low-level skills (e.g., standing, walking, and running) that can then be re-used to synthesize high-level behaviors. However, even with MoCap data, controlling simulated humanoids remains very hard, as MoCap data offers only kinematic information. Finding physical control inputs to realize the demonstrated motions requires computationally intensive methods like reinforcement learning. Thus, despite the publicly available MoCap data, its utility has been limited to institutions with large-scale compute. In this work, we dramatically lower the barrier for productive research on this topic by training and releasing high-quality agents that can track over three hours of MoCap data for a simulated humanoid in the dmcontrol physics-based environment. We release MoCapAct (Motion Capture with Actions), a dataset of these expert agents and their rollouts, which contain proprioceptive observations and actions. We demonstrate the utility of MoCapAct by using it to train a single hierarchical policy capable of tracking the entire MoCap dataset within dmcontrol and show the learned low-level component can be re-used to efficiently learn downstream high-level tasks. Finally, we use MoCapAct to train an autoregressive GPT model and show that it can control a simulated humanoid to perform natural motion completion given a motion prompt.Videos of the results and links to the code and dataset are available at https://microsoft.github.io/MoCapAct.

We provide a framework to accelerate reinforcement learning (RL) algorithms by heuristics that are constructed by domain knowledge or offline data. Tabula rasa RL algorithms require environment interactions or computation that scales with the horizon of the sequential decision-making task. Using our framework, we show how heuristic-guided RL induces a much shorter horizon sub-problem that provably solves the original task. Our framework can be viewed as a horizon-based regularization for controlling bias and variance in RL under a finite interaction budget. In theory, we characterize the properties of a good heuristic and the resulting impact on RL acceleration. In particular, we introduce the novel concept of an improvable heuristic that can allow any RL agent to conservatively extrapolate beyond its prior knowledge. In practice, we instantiate our framework to accelerate several state-of-the-art algorithms in simulated robotic control tasks and procedurally generated games. Our framework complements the rich literature on warm-starting RL using expert demonstrations or exploratory data-sets, and creates a unified channel to inject prior knowledge into RL.

Existing multi-armed bandit (MAB) models make two implicit assumptions: an arm generates a payoff only when it is played, and the agent observes every payoff that is generated. This paper introduces synchronization bandits, a MAB variant where all arms generate costs at all times, but the agent observes an arm’s instantaneous cost only when the arm is played. Synchronization MABs are inspired by online caching scenarios such as Web crawling, where an arm corresponds to a cached item and playing the arm means downloading its fresh copy from a server. We present MirrorSync, an online learning algorithm for synchronization bandits, establish an adversarial regret of $O(T^{2/3})$ for it, and show how to make it practical.

Despite its promise, reinforcement learning’s real-world adoption has been hampered by the need for costly exploration to learn a good policy. Imitation learning (IL) mitigates this shortcoming by using an oracle policy during training as a bootstrap to accelerate the learning process. However, in many practical situations, the learner has access to multiple suboptimal oracles, which may provide conflicting advice in a state. The existing IL literature provides a limited treatment of such scenarios. Whereas in the single-oracle case, the return of the oracle’s policy provides an obvious benchmark for the learner to compete against, neither such a benchmark nor principled ways of outperforming it are known for the multi-oracle setting. In this paper, we propose the state-wise maximum of the oracle policies’ values as a natural baseline to resolve conflicting advice from multiple oracles. Using a reduction of policy optimization to online learning, we introduce a novel IL algorithm MAMBA, which can provably learn a policy competitive with this benchmark. In particular, MAMBA optimizes policies by using a gradient estimator in the style of generalized advantage estimation (GAE). Our theoretical analysis shows that this design makes MAMBA robust and enables it to outperform the oracle policies by a larger margin than the IL state of the art, even in the single-oracle case. In an evaluation against standard policy gradient with GAE and AggreVaTe(D), we showcase MAMBA’s ability to leverage demonstrations both from a single and from multiple weak oracles, and significantly speed up policy optimization.

In safety-critical applications, autonomous agents may need to learn in an environment where mistakes can be very costly. In such settings, the agent needs to behave safely not only after but also while learning. To achieve this, existing safe reinforcement learning methods make an agent rely on priors that let it avoid dangerous situations during exploration with high probability, but both the probabilistic guarantees and the smoothness assumptions inherent in the priors are not viable in many scenarios of interest such as autonomous driving. This paper presents an alternative approach inspired by human teaching, where an agent learns under the supervision of an automatic instructor that saves the agent from violating constraints during learning. In this model, we introduce the monitor that neither needs to know how to do well at the task the agent is learning nor needs to know how the environment works. Instead, it has a library of reset controllers that it activates when the agent starts behaving dangerously, preventing it from doing damage. Crucially, the choices of which reset controller to apply in which situation affect the speed of agent learning. Based on observing agents' progress the teacher itself learns a policy for choosing the reset controllers, a curriculum, to optimize the agent's final policy reward. Our experiments use this framework in two environments to induce curricula for safe and efficient learning.

From traditional Web search engines to virtual assistants and Web accelerators, services that rely on online information need to continually keep track of remote content changes by explicitly requesting content updates from remote sources (e.g., web pages). We propose a novel optimization objective for this setting that has several practically desirable properties, and efficient algorithms for it with optimality guarantees even in the face of mixed content change observability and initially unknown change model parameters. Experiments on 18.5M URLs crawled daily for 14 weeks show significant advantages of this approach over prior art.

A Web crawler is an essential part of a search engine that procures information subsequently served by the search engine to its users. As the Web is becoming increasingly more dynamic, in addition to discovering new web pages a crawler needs to keep revisiting those already in the search engine's index, in order to keep the index fresh by picking up the pages' changed content. Determining how often to recrawl pages requires making tradeoffs based on the pages' relative importance and change rates, subject to multiple resource constraints - the limited daily budget of crawl requests on the search engine's end and politeness constraints restricting the rate at which pages can be requested from a given host. In this paper, we introduce PoliteBinaryLambdaCrawl, the first optimal algorithm for freshness crawl scheduling in the presence of politeness constraints as well as non-uniform page importance scores and the crawler's own crawl request limit. We also propose an approximation for it, stating its theoretical optimality conditions and in the process discovering a connection to an approach previously thought of as a mere heuristic for freshness crawl scheduling. We explore the relative performance of PoliteBinaryLambdaCrawl and other methods for handling politeness constraints on a dataset collected by crawling over 18.5M URLs daily over 14 weeks.

Autonomous soaring capability has the potential to significantly increase time aloft for fixed-wing UAVs. In this paper, we introduce ArduSoar, the first soaring controller integrated into a major autopilot software suite for small UAVs. We describe ArduSoar from the algorithmic standpoint, outline its integration with the ArduPlane autopilot, discuss parameter tuning for it, and conduct a series of flight tests on real sUAVs that show ArduSoar's robustness even in highly nonideal atmospheric conditions.

The traditional way of obtaining models from data, inductive learning, has proved itself both in theory and in many practical applications. However, in domains where data is difficult or expensive to obtain, e.g., medicine, deep transfer learning is a more promising technique. It circumvents the model acquisition difficulties caused by scarce data in a target domain by carrying over structural properties of a model learned in a source domain where training data is ample. Nonetheless, the lack of a principled view of transfer learning so far has limited its adoption. In this paper, we address this issue by regarding transfer learning as a process that biases learning in a target domain in favor of patterns useful in a source domain. Specifically, we consider a first-order logic model of the data as an instantiation of a set of second-order templates. Hence, the usefulness of a model is partly determined by the learner's prior distribution over these template sets. The main insight of our work is that transferring knowledge amounts to acquiring a posterior over the second-order template sets by learning in the source domain and using this posterior when learning in the target setting. Our experimental evaluation demonstrates our approach to outperform the existing transfer learning techniques in terms of accuracy and runtime.

In many probabilistic planning scenarios, a system’s behavior needs to not only maximize the expected utility but also obey certain restrictions. This paper presents Saturated Path-Constrained Markov Decision Processes (SPC MDPs), a new MDP type for planning under uncertainty with deterministic model-checking constraints, e.g., "state s must be visited befores s'", "the system must end up in s", or "the system must never enter s". We present a mathematical analysis of SPCMDPs, showing that although SPC MDPs generally have no optimal policies, every instance of this class has an epsilon-optimal randomized policy for any > 0. We propose a dynamic programming-based algorithm for finding such policies, and empirically demonstrate this algorithm to be orders of magnitude faster than its next-best alternative.

An ideal crowdsourcing or citizen-science system would route tasks to the most appropriate workers, but the best assignment is unclear because workers have varying skill, tasks have varying difficulty, and assigning several workers to a single task may significantly improve output quality. This paper defines a space of task routing problems, proves that even the simplest is NP-hard, and develops several approximation algorithms for parallel routing problems. We show that an intuitive class of requesters' utility functions is submodular, which lets us provide iterative methods for dynamically allocating batches of tasks that make near-optimal use of available workers in each round. Experiments with live oDesk workers show that our task routing algorithm uses only 48% of the human labor compared to the commonly used round-robin strategy. Further, we provide versions of our task routing algorithm which enable it to scale to large numbers of workers and questions and to handle workers with variable response times while still providing significant benefit over common baselines.

UCT, the premier method for solving games such as Go, is also becoming the dominant algorithm for probabilistic planning. Out of the five solvers at the International Probabilistic Planning Competition (IPPC) 2011, four were based on the UCT algorithm. However, while a UCT-based planner, PROST, won the contest, an LRTDP-based system, Glutton, came in a close second, outperforming other systems derived from UCT. These results raise a question: what are the strengths and weaknesses of LRTDP and UCT in practice? This paper starts answering this question by contrasting the two approaches in the context of finite-horizon MDPs. We demonstrate that in such scenarios, UCT's lack of a sound termination condition is a serious practical disadvantage. In order to handle an MDP with a large finite horizon under a time constraint, UCT forces an expert to guess a non-myopic lookahead value for which it should be able to converge on the encountered states. Mistakes in setting this parameter can greatly hurt UCT's performance. In contrast, LRTDP's convergence criterion allows for an iterative deepening strategy. Using this strategy, LRTDP automatically finds the largest lookahead value feasible under the given time constraint. As a result, LRTDP has better performance and stronger theoretical properties. We present an online version of Glutton, named Gourmand, that illustrates this analysis and outperforms PROST on the set of IPPC-2011 problems.

The results of the latest International Probabilistic Planning Competition (IPPC-2008) indicate that the presence of dead ends, states with no trajectory to the goal, makes MDPs hard for modern probabilistic planners. Implicit dead ends, states with executable actions but no path to the goal, are particularly challenging; existing MDP solvers spend much time and memory identifying these states. As a first attempt to address this issue, we propose a machine learning algorithm called SIXTHSENSE. SIXTHSENSE helps existing MDP solvers by finding nogoods, conjunctions of literals whose truth in a state implies that the state is a dead end. Importantly, our learned nogoods are sound, and hence the states they identify are true dead ends. SIXTHSENSE is very fast, needs little training data, and takes only a small fraction of total planning time. While IPPC problems may have millions of dead ends, they may typically be represented with only a dozen or two no-goods. Thus, nogood learning efficiently produces a quick and reliable means for dead-end recognition. Our experiments show that the nogoods found by SIXTHSENSE routinely reduce planning space and time on IPPC domains, enabling some planners to solve problems they could not previously handle.