Not yet. We present SPACE, a benchmark that systematically evaluates spatial cognition in frontier models. Our benchmark builds on decades of research in cognitive science. It evaluates large-scale mapping abilities that are brought to bear when an organism traverses physical environments, smaller-scale reasoning about object shapes and layouts, and cognitive infrastructure such as spatial attention and memory. For many tasks, we instantiate parallel presentations via text and images, allowing us to benchmark both large language models and large multimodal models. Results suggest that contemporary frontier models fall short of the spatial intelligence of animals, performing near chance level on a number of classic tests of animal cognition.

As language models grow ever larger, so do their vocabularies. This has shifted the memory footprint of LLMs during training disproportionately to one single layer: the cross-entropy in the loss computation. Cross-entropy builds up a logit matrix with entries for each pair of input tokens and vocabulary items and, for small models, consumes an order of magnitude more memory than the rest of the LLM combined. We propose Cut Cross-Entropy (CCE), a method that computes the cross-entropy loss without materializing the logits for all tokens into global memory. Rather, CCE only computes the logit for the correct token and evaluates the log-sum-exp over all logits on the fly. We implement a custom kernel that performs the matrix multiplications and the log-sum-exp reduction over the vocabulary in flash memory, making global memory consumption for the cross-entropy computation negligible. This has a dramatic effect. Taking the Gemma 2 (2B) model as an example, CCE reduces the memory footprint of the loss computation from 24 GB to 1 MB, and the total training-time memory consumption of the classifier head from 28 GB to 1 GB. To improve the throughput of CCE, we leverage the inherent sparsity of softmax and propose to skip elements of the gradient computation that have a negligible (i.e. below numerical precision) contribution to the gradient. Experiments demonstrate that the dramatic reduction in memory consumption is accomplished without sacrificing training speed or convergence.

Self-play has powered breakthroughs in two-player and multi-player games. Here we show that self-play is a surprisingly effective strategy in another domain. We show that robust and naturalistic driving emerges entirely from self-play in simulation at unprecedented scale -- 1.6 billion km of driving. This is enabled by Gigaflow, a batched simulator that can synthesize and train on 42 years of subjective driving experience per hour on a single 8-GPU node. The resulting policy achieves state-of-the-art performance on three independent autonomous driving benchmarks. The policy outperforms the prior state of the art when tested on recorded real-world scenarios, amidst human drivers, without ever seeing human data during training. The policy is realistic when assessed against human references and achieves unprecedented robustness, averaging 17.5 years of continuous driving between incidents in simulation.

We study ObjectGoal Navigation - where a virtual robot situated in a new environment is asked to navigate to an object. Prior work [1] has shown that imitation learning (IL) using behavior cloning (BC) on a dataset of human demonstrations achieves promising results. However, this has limitations - 1) BC policies generalize poorly to new states, since the training mimics actions not their consequences, and 2) collecting demonstrations is expensive. On the other hand, reinforcement learning (RL) is trivially scalable, but requires careful reward engineering to achieve desirable behavior. We present PIRLNav, a two-stage learning scheme for BC pretraining on human demonstrations followed by RL-finetuning. This leads to a policy that achieves a success rate of 65.0% on OBJECTNAV (+5.0% absolute over previous state-of-the-art). Using this BC→RL training recipe, we present a rigorous empirical analysis of design choices. First, we investigate whether human demonstrations can be replaced with ‘free’ (automatically generated) sources of demonstrations, e.g. shortest paths (SP) or task-agnostic frontier exploration (FE) trajectories. We find that BC→RL on human demonstrations outperforms BC→RL on SP and FE trajectories, even when controlled for the same BC-pretraining success on TRAIN, and even on a subset of VAL episodes where BC-pretraining success favors the SP or FE policies. Next, we study how RL-finetuning performance scales with the size of the BC pretraining dataset. We find that as we increase the size of the BC-pretraining dataset and get to high BC accuracies, the improvements from RL-finetuning are smaller, and that 90% of the performance of our best BC→RL policy can be achieved with less than half the number of BC demonstrations. Finally, we analyze failure modes of our OBJECTNAV policies, and present guidelines for further improving them. Project page: ram81.github.io/projects/pirlnav.

Training AI agents to perform complex tasks in simulated worlds requires millions to billions of steps of experience. To achieve high performance, today's fastest simulators for training AI agents adopt the idea of batch simulation: using a single simulation engine to simultaneously step many environments in parallel. We introduce a framework for productively authoring novel training environments (including custom logic for environment generation, environment time stepping, and generating agent observations and rewards) that execute as high-performance, GPU-accelerated batched simulators. Our key observation is that the entity-component-system (ECS) design pattern, popular for expressing CPU-side game logic today, is also well-suited for providing the structure needed for high-performance batched simulators. We contribute the first fully-GPU accelerated ECS implementation that natively supports batch environment simulation. We demonstrate how ECS abstractions impose structure on a training environment's logic and state that allows the system to efficiently manage state, amortize work, and identify GPU-friendly coherent parallel computations within and across different environments. We implement several learning environments in this framework, and demonstrate GPU speedups of two to three orders of magnitude over open source CPU baselines and 5-33× over strong baselines running on a 32-thread CPU. An implementation of the OpenAI hide and seek 3D environment written in our framework, which performs rigid body physics and ray tracing in each simulator step, achieves over 1.9 million environment steps per second on a single GPU.

Can an autonomous agent navigate in a new environment without building an explicit map? For the task of PointGoal navigation ('Go to Δx, Δy’) under idealized settings (no RGB-D and actuation noise, perfect GPS+Compass), the answer is a clear ‘yes' - mapless neural models composed of task-agnostic components (CNNs and RNNs) trained with large-scale reinforcement learning achieve 100% Success on a standard dataset (Gibson [24] ). However, for PointNav in a realistic setting (RGB-D and actuation noise, no GPS+Compass), this is an open question; one we tackle in this paper. The strongest published result for this task is 71.7% Success [39].11According to Habitat Challenge 2020 PointNav benchmark held annually. A concurrent as-yet-unpublished result has reported 91% Success on 2021's benchmark, but we are unable to comment on the details because an associated report is not available.First, we identify the main (perhaps, only) cause of the drop in performance: absence of GPS+Compass. An agent with perfect GPS+Compass faced with RGB-D sensing and actuation noise achieves 99.8% Success (Gibson- v2 val). This suggests that (to paraphrase a meme) robust visual odometry is all we need for realistic PointNav; if we can achieve that, we can ignore the sensing and actuation noise. With that as our operating hypothesis, we scale dataset size, model size, and develop human-annotation-free dataaugmentation techniques to train neural models for visual odometry. We advance state of the art on the Habitat Realistic PointNav Challenge - SPL by 40% (relative), 53 to 74, and Success by 31% (relative), 71 to 94. While our approach does not saturate or ‘solve’ this dataset, this strong improvement combined with promising zero-shot sim2real transfer (to a LoCoBot robot) provides evidence consistent with the hypothesis that explicit mapping may not be necessary for navigation, even in a realistic setting.

We present Variable Experience Rollout (VER), a technique for efficiently scaling batched on-policy reinforcement learning in heterogenous environments (where different environments take vastly different times to generate rollouts) to many GPUs residing on, potentially, many machines. VER combines the strengths of and blurs the line between synchronous and asynchronous on-policy RL methods (SyncOnRL and AsyncOnRL, respectively). Specifically, it learns from on-policy experience (like SyncOnRL) and has no synchronization points (like AsyncOnRL) enabling high throughput.We find that VER leads to significant and consistent speed-ups across a broad range of embodied navigation and mobile manipulation tasks in photorealistic 3D simulation environments. Specifically, for PointGoal navigation and ObjectGoal navigation in Habitat 1.0, VER is 60-100% faster (1.6-2x speedup) than DD-PPO, the current state of art for distributed SyncOnRL, with similar sample efficiency. For mobile manipulation tasks (open fridge/cabinet, pick/place objects) in Habitat 2.0 VER is 150% faster (2.5x speedup) on 1 GPU and 170% faster (2.7x speedup) on 8 GPUs than DD-PPO. Compared to SampleFactory (the current state-of-the-art AsyncOnRL), VER matches its speed on 1 GPU, and is 70% faster (1.7x speedup) on 8 GPUs with better sample efficiency.We leverage these speed-ups to train chained skills for GeometricGoal rearrangement tasks in the Home Assistant Benchmark (HAB). We find a surprising emergence of navigation in skills that do not ostensible require any navigation. Specifically, the Pick skill involves a robot picking an object from a table. During training the robot was always spawned close to the table and never needed to navigate. However, we find that if base movement is part of the action space, the robot learns to navigate then pick an object in new environments with 50% success, demonstrating surprisingly high out-of-distribution generalization.

Can general-purpose neural models learn to navigate? For PointGoal navigation (‘go to Δx, Δy’), the answer is a clear ‘yes’ – mapless neural models composed of task-agnostic components (CNNs and RNNs) trained with large-scale model-free reinforcement learning achieve near-perfect performance [27]. However, for ObjectGoal navigation (‘find a TV’), this is an open question; one we tackle in this paper. The current best-known result on ObjectNav with general-purpose models is 6% success rate [25].First, we show that the key problem is overfitting. Large-scale training results in 94% success rate on training environments and only 8% in validation. We observe that this stems from agents memorizing environment layouts during training – sidestepping the need for exploration and directly learning shortest paths to nearby goal objects. We show that this is a natural consequence of optimizing for the task metric (which in fact penalizes exploration), is enabled by powerful observation encoders, and is possible due to the finite set of training environment configurations.Informed by our findings, we introduce Treasure Hunt Data Augmentation (THDA) to address overfitting in ObjectNav. THDA inserts 3D scans of household objects at arbitrary scene locations and uses them as ObjectNav goals – augmenting and greatly expanding the set of training layouts. Taken together with our other proposed changes, we improve the state of art on the Habitat ObjectGoal Navigation benchmark by 90% (from 14% success rate to 27%) and path efficiency by 48% (from 7.5 SPL to 11.1 SPL).

ObjectGoal Navigation (ObjectNav) is an embodied task wherein agents are to navigate to an object instance in an unseen environment. Prior works have shown that end-to-end ObjectNav agents that use vanilla visual and recurrent modules, e.g. a CNN+RNN, perform poorly due to overfitting and sample inefficiency. This has motivated current state-of-the-art methods to mix analytic and learned components and operate on explicit spatial maps of the environment. We instead re-enable a generic learned agent by adding auxiliary learning tasks and an exploration reward. Our agents achieve 24.5% success and 8.1% SPL, a 37% and 8% relative improvement over prior state-of-the-art, respectively, on the Habitat ObjectNav Challenge [35]. From our analysis, we propose that agents will act to simplify their visual inputs so as to smooth their RNN dynamics, and that auxiliary tasks reduce overfitting by minimizing effective RNN dimensionality; i.e. a performant ObjectNav agent that must maintain coherent plans over long horizons does so by learning smooth, low-dimensional recurrent dynamics. Site: joel99.github.io/objectnav/

We present Megaverse, a new 3D simulation platform for reinforcement learning and embodied AI research. The efficient design of our engine enables physics-based simulation with high-dimensional egocentric observations at more than 1,000,000 actions per second on a single 8-GPU node. Megaverse is up to 70x faster than DeepMind Lab in fully-shaded 3D scenes with interactive objects. We achieve this high simulation performance by leveraging batched simulation, thereby taking full advantage of the massive parallelism of modern GPUs. We use Megaverse to build a new benchmark that consists of several single-agent and multi-agent tasks covering a variety of cognitive challenges. We evaluate model-free RL on this benchmark to provide baselines and facilitate future research.

We introduce Habitat 2.0 (H2.0), a simulation platform for training virtual robots in interactive 3D environments and complex physics-enabled scenarios. We make comprehensive contributions to all levels of the embodied AI stack – data, simulation, and benchmark tasks. Specifically, we present: (i) ReplicaCAD: an artist-authored, annotated, reconfigurable 3D dataset of apartments (matching real spaces) with articulated objects (e.g. cabinets and drawers that can open/close); (ii) H2.0: a high-performance physics-enabled 3D simulator with speeds exceeding 25,000 simulation steps per second (850x real-time) on an 8-GPU node, representing 100x speed-ups over prior work; and, (iii) Home Assistant Benchmark (HAB): a suite of common tasks for assistive robots (tidy the house, stock groceries, set the table) that test a range of mobile manipulation capabilities. These large-scale engineering contributions allow us to systematically compare deep reinforcement learning (RL) at scale and classical sense-plan-act (SPA) pipelines in long-horizon structured tasks, with an emphasis on generalization to new objects, receptacles, and layouts. We find that (1) flat RL policies struggle on HAB compared to hierarchical ones; (2) a hierarchy with independent skills suffers from ‘hand-off problems’, and (3) SPA pipelines are more brittle than RL policies.

To help bridge the gap between internet vision-style problems and the goal of vision for embodied perception we instantiate a large-scale navigation task -- Embodied Question Answering [1] in photo-realistic environments (Matterport 3D). We thoroughly study navigation policies that utilize 3D point clouds, RGB images, or their combination. Our analysis of these models reveals several key findings. We find that two seemingly naive navigation baselines, forward-only and random, are strong navigators and challenging to outperform, due to the specific choice of the evaluation setting presented by [1]. We find a novel loss-weighting scheme we call Inflection Weighting to be important when training recurrent models for navigation with behavior cloning and are able to out perform the baselines with this technique. We find that point clouds provide a richer signal than RGB images for learning obstacle avoidance, motivating the use (and continued study) of 3D deep learning models for embodied navigation.

We present Habitat, a platform for research in embodied artificial intelligence (AI). Habitat enables training embodied agents (virtual robots) in highly efficient photorealistic 3D simulation. Specifically, Habitat consists of: (i) Habitat-Sim: a flexible, high-performance 3D simulator with configurable agents, sensors, and generic 3D dataset handling. Habitat-Sim is fast - when rendering a scene from Matterport3D, it achieves several thousand frames per second (fps) running single-threaded, and can reach over 10,000 fps multi-process on a single GPU. (ii) Habitat-API: a modular high-level library for end-toend development of embodied AI algorithms - defining tasks (e.g. navigation, instruction following, question answering), configuring, training, and benchmarking embodied agents. These large-scale engineering contributions enable us to answer scientific questions requiring experiments that were till now impracticable or `merely' impractical. Specifically, in the context of point-goal navigation: (1) we revisit the comparison between learning and SLAM approaches from two recent works [19, 16] and find evidence for the opposite conclusion - that learning outperforms SLAM if scaled to an order of magnitude more experience than previous investigations, and (2) we conduct the first cross-dataset generalization experiments {train, test} × {Matterport3D, Gibson} for multiple sensors {blind, RGB, RGBD, D} and find that only agents with depth (D) sensors generalize across datasets. We hope that our open-source platform and these findings will advance research in embodied AI.

This paper presents a novel algorithm that utilizes a 2D floorplan to align panorama RGBD scans. While effective panorama RGBD alignment techniques exist, such a system requires extremely dense RGBD image sampling. Our approach can significantly reduce the number of necessary scans with the aid of a floorplan image. We formulate a novel Markov Random Field inference problem as a scan placement over the floorplan, as opposed to the conventional scan-to-scan alignment. The technical contributions lie in multi-modal image correspondence cues (between scans and schematic floorplan) as well as a novel coverage potential avoiding an inherent stacking bias. The proposed approach has been evaluated on five challenging large indoor spaces. To the best of our knowledge, we present the first effective system that utilizes a 2D floorplan image for building-scale 3D pointcloud alignment. The source code and the data are shared with the community to further enhance indoor mapping research.