arXiv subjects

Erik Wijmans

Publications and source records attributed to Erik Wijmans.

At least 19 recordsLinked to original sources

arXiv2026

Entropy-Preserving Reinforcement Learning

Policy gradient algorithms have driven many recent advancements in language model reasoning. An appealing property is their ability to learn from exploration on their own trajectories, a process crucial for fostering diverse and creative solutions. As we show in this paper, many policy gradient algorithms naturally reduce the entropy -- and thus the diversity of explored trajectories -- as part of training, yielding a policy increasingly limited in its ability to explore. In this paper, we argue that entropy should be actively monitored and controlled throughout training. We formally analyze the contributions of leading policy gradient objectives on entropy dynamics, identify empirical factors (such as numerical precision) that significantly impact entropy behavior, and propose explicit mechanisms for entropy control. These include REPO, a family of algorithms that modify the advantage function to regulate entropy, and ADAPO, an adaptive asymmetric clipping approach. Models trained with our entropy-preserving methods maintain diversity throughout training, yielding final policies that are more performant and retain their trainability for sequential learning in new environments.

cs.LG↗

arXiv2025

Does Spatial Cognition Emerge in Frontier Models?

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. Code and data are available: https://github.com/apple/ml-space-benchmark

cs.AI↗

arXiv2025

Cut Your Losses in Large-Vocabulary Language Models

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.

cs.LG↗

arXiv2025

Robust Autonomy Emerges from Self-Play

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.

cs.LG↗

arXiv2023

PIRLNav: Pretraining with Imitation and RL Finetuning for ObjectNav

We study ObjectGoal Navigation -- where a virtual robot situated in a new environment is asked to navigate to an object. Prior work 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$\rightarrow$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$\rightarrow$RL on human demonstrations outperforms BC$\rightarrow$RL on SP and FE trajectories, even when controlled for 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 BC-pretraining dataset and get to high BC accuracies, improvements from RL-finetuning are smaller, and that $90\%$ of the performance of our best BC$\rightarrow$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.

cs.LG↗

arXiv2023

Emergence of Maps in the Memories of Blind Navigation Agents

Animal navigation research posits that organisms build and maintain internal spatial representations, or maps, of their environment. We ask if machines -- specifically, artificial intelligence (AI) navigation agents -- also build implicit (or 'mental') maps. A positive answer to this question would (a) explain the surprising phenomenon in recent literature of ostensibly map-free neural-networks achieving strong performance, and (b) strengthen the evidence of mapping as a fundamental mechanism for navigation by intelligent embodied agents, whether they be biological or artificial. Unlike animal navigation, we can judiciously design the agent's perceptual system and control the learning paradigm to nullify alternative navigation mechanisms. Specifically, we train 'blind' agents -- with sensing limited to only egomotion and no other sensing of any kind -- to perform PointGoal navigation ('go to $Δ$ x, $Δ$ y') via reinforcement learning. Our agents are composed of navigation-agnostic components (fully-connected and recurrent neural networks), and our experimental setup provides no inductive bias towards mapping. Despite these harsh conditions, we find that blind agents are (1) surprisingly effective navigators in new environments (~95% success); (2) they utilize memory over long horizons (remembering ~1,000 steps of past experience in an episode); (3) this memory enables them to exhibit intelligent behavior (following walls, detecting collisions, taking shortcuts); (4) there is emergence of maps and collision detection neurons in the representations of the environment built by a blind agent as it navigates; and (5) the emergent maps are selective and task dependent (e.g. the agent 'forgets' exploratory detours). Overall, this paper presents no new techniques for the AI audience, but a surprising finding, an insight, and an explanation.

cs.AI↗

arXiv2022

VER: Scaling On-Policy RL Leads to the Emergence of Navigation in Embodied Rearrangement

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). VER learns from on-policy experience (like SyncOnRL) and has no synchronization points (like AsyncOnRL). 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 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.

cs.LG↗

arXiv2022

Habitat 2.0: Training Home Assistants to Rearrange their Habitat

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, prepare 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.

cs.LG↗

arXiv2022

Is Mapping Necessary for Realistic PointGoal Navigation?

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' - map-less neural models composed of task-agnostic components (CNNs and RNNs) trained with large-scale reinforcement learning achieve 100% Success on a standard dataset (Gibson). 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. First, we identify the main (perhaps, only) cause of the drop in performance: the 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 the dataset and model size, and develop human-annotation-free data-augmentation techniques to train models for visual odometry. We advance the state of art on the Habitat Realistic PointNav Challenge from 71% to 94% Success (+23, 31% relative) and 53% to 74% SPL (+21, 40% relative). While our approach does not saturate or 'solve' this dataset, this strong improvement combined with promising zero-shot sim2real transfer (to a LoCoBot) provides evidence consistent with the hypothesis that explicit mapping may not be necessary for navigation, even in a realistic setting.

cs.CV↗

arXiv2021

Realistic PointGoal Navigation via Auxiliary Losses and Information Bottleneck

We propose a novel architecture and training paradigm for training realistic PointGoal Navigation -- navigating to a target coordinate in an unseen environment under actuation and sensor noise without access to ground-truth localization. Specifically, we find that the primary challenge under this setting is learning localization -- when stripped of idealized localization, agents fail to stop precisely at the goal despite reliably making progress towards it. To address this we introduce a set of auxiliary losses to help the agent learn localization. Further, we explore the idea of treating the precise location of the agent as privileged information -- it is unavailable during test time, however, it is available during training time in simulation. We grant the agent restricted access to ground-truth localization readings during training via an information bottleneck. Under this setting, the agent incurs a penalty for using this privileged information, encouraging the agent to only leverage this information when it is crucial to learning. This enables the agent to first learn navigation and then learn localization instead of conflating these two objectives in training. We evaluate our proposed method both in a semi-idealized (noiseless simulation without Compass+GPS) and realistic (addition of noisy simulation) settings. Specifically, our method outperforms existing baselines on the semi-idealized setting by 18\%/21\% SPL/Success and by 15\%/20\% SPL in the realistic setting. Our improved Success and SPL metrics indicate our agent's improved ability to accurately self-localize while maintaining a strong navigation policy. Our implementation can be found at https://github.com/NicoGrande/habitat-pointnav-via-ib.

cs.CV↗

arXiv2021

Habitat-Matterport 3D Dataset (HM3D): 1000 Large-scale 3D Environments for Embodied AI

We present the Habitat-Matterport 3D (HM3D) dataset. HM3D is a large-scale dataset of 1,000 building-scale 3D reconstructions from a diverse set of real-world locations. Each scene in the dataset consists of a textured 3D mesh reconstruction of interiors such as multi-floor residences, stores, and other private indoor spaces. HM3D surpasses existing datasets available for academic research in terms of physical scale, completeness of the reconstruction, and visual fidelity. HM3D contains 112.5k m^2 of navigable space, which is 1.4 - 3.7x larger than other building-scale datasets such as MP3D and Gibson. When compared to existing photorealistic 3D datasets such as Replica, MP3D, Gibson, and ScanNet, images rendered from HM3D have 20 - 85% higher visual fidelity w.r.t. counterpart images captured with real cameras, and HM3D meshes have 34 - 91% fewer artifacts due to incomplete surface reconstruction. The increased scale, fidelity, and diversity of HM3D directly impacts the performance of embodied AI agents trained using it. In fact, we find that HM3D is `pareto optimal' in the following sense -- agents trained to perform PointGoal navigation on HM3D achieve the highest performance regardless of whether they are evaluated on HM3D, Gibson, or MP3D. No similar claim can be made about training on other datasets. HM3D-trained PointNav agents achieve 100% performance on Gibson-test dataset, suggesting that it might be time to retire that episode dataset.

cs.CV↗

arXiv2021

Auxiliary Tasks and Exploration Enable ObjectNav

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. 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: https://joel99.github.io/objectnav/

cs.CV↗

arXiv2021

Megaverse: Simulating Embodied Agents at One Million Experiences per Second

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. The source code is available at https://www.megaverse.info

cs.LG↗

arXiv2021

Large Batch Simulation for Deep Reinforcement Learning

We accelerate deep reinforcement learning-based training in visually complex 3D environments by two orders of magnitude over prior work, realizing end-to-end training speeds of over 19,000 frames of experience per second on a single GPU and up to 72,000 frames per second on a single eight-GPU machine. The key idea of our approach is to design a 3D renderer and embodied navigation simulator around the principle of "batch simulation": accepting and executing large batches of requests simultaneously. Beyond exposing large amounts of work at once, batch simulation allows implementations to amortize in-memory storage of scene assets, rendering work, data loading, and synchronization costs across many simulation requests, dramatically improving the number of simulated agents per GPU and overall simulation throughput. To balance DNN inference and training costs with faster simulation, we also build a computationally efficient policy DNN that maintains high task performance, and modify training algorithms to maintain sample efficiency when training with large mini-batches. By combining batch simulation and DNN performance optimizations, we demonstrate that PointGoal navigation agents can be trained in complex 3D environments on a single GPU in 1.5 days to 97% of the accuracy of agents trained on a prior state-of-the-art system using a 64-GPU cluster over three days. We provide open-source reference implementations of our batch 3D renderer and simulator to facilitate incorporation of these ideas into RL systems.

cs.LG↗

arXiv2020

How to Train PointGoal Navigation Agents on a (Sample and Compute) Budget

PointGoal navigation has seen significant recent interest and progress, spurred on by the Habitat platform and associated challenge. In this paper, we study PointGoal navigation under both a sample budget (75 million frames) and a compute budget (1 GPU for 1 day). We conduct an extensive set of experiments, cumulatively totaling over 50,000 GPU-hours, that let us identify and discuss a number of ostensibly minor but significant design choices -- the advantage estimation procedure (a key component in training), visual encoder architecture, and a seemingly minor hyper-parameter change. Overall, these design choices to lead considerable and consistent improvements over the baselines present in Savva et al. Under a sample budget, performance for RGB-D agents improves 8 SPL on Gibson (14% relative improvement) and 20 SPL on Matterport3D (38% relative improvement). Under a compute budget, performance for RGB-D agents improves by 19 SPL on Gibson (32% relative improvement) and 35 SPL on Matterport3D (220% relative improvement). We hope our findings and recommendations will make serve to make the community's experiments more efficient.

cs.CV↗

arXiv2020

Auxiliary Tasks Speed Up Learning PointGoal Navigation

PointGoal Navigation is an embodied task that requires agents to navigate to a specified point in an unseen environment. Wijmans et al. showed that this task is solvable but their method is computationally prohibitive, requiring 2.5 billion frames and 180 GPU-days. In this work, we develop a method to significantly increase sample and time efficiency in learning PointNav using self-supervised auxiliary tasks (e.g. predicting the action taken between two egocentric observations, predicting the distance between two observations from a trajectory,etc.).We find that naively combining multiple auxiliary tasks improves sample efficiency,but only provides marginal gains beyond a point. To overcome this, we use attention to combine representations learnt from individual auxiliary tasks. Our best agent is 5.5x faster to reach the performance of the previous state-of-the-art, DD-PPO, at 40M frames, and improves on DD-PPO's performance at 40M frames by 0.16 SPL. Our code is publicly available at https://github.com/joel99/habitat-pointnav-aux.

cs.CV↗

arXiv2020

ObjectNav Revisited: On Evaluation of Embodied Agents Navigating to Objects

We revisit the problem of Object-Goal Navigation (ObjectNav). In its simplest form, ObjectNav is defined as the task of navigating to an object, specified by its label, in an unexplored environment. In particular, the agent is initialized at a random location and pose in an environment and asked to find an instance of an object category, e.g., find a chair, by navigating to it. As the community begins to show increased interest in semantic goal specification for navigation tasks, a number of different often-inconsistent interpretations of this task are emerging. This document summarizes the consensus recommendations of this working group on ObjectNav. In particular, we make recommendations on subtle but important details of evaluation criteria (for measuring success when navigating towards a target object), the agent's embodiment parameters, and the characteristics of the environments within which the task is carried out. Finally, we provide a detailed description of the instantiation of these recommendations in challenges organized at the Embodied AI workshop at CVPR 2020 http://embodied-ai.org .

cs.CV↗

arXiv2020

Sim2Real Predictivity: Does Evaluation in Simulation Predict Real-World Performance?

Does progress in simulation translate to progress on robots? If one method outperforms another in simulation, how likely is that trend to hold in reality on a robot? We examine this question for embodied PointGoal navigation, developing engineering tools and a research paradigm for evaluating a simulator by its sim2real predictivity. First, we develop Habitat-PyRobot Bridge (HaPy), a library for seamless execution of identical code on simulated agents and robots, transferring simulation-trained agents to a LoCoBot platform with a one-line code change. Second, we investigate the sim2real predictivity of Habitat-Sim for PointGoal navigation. We 3D-scan a physical lab space to create a virtualized replica, and run parallel tests of 9 different models in reality and simulation. We present a new metric called Sim-vs-Real Correlation Coefficient (SRCC) to quantify predictivity. We find that SRCC for Habitat as used for the CVPR19 challenge is low (0.18 for the success metric), suggesting that performance differences in this simulator-based challenge do not persist after physical deployment. This gap is largely due to AI agents learning to exploit simulator imperfections, abusing collision dynamics to 'slide' along walls, leading to shortcuts through otherwise non-navigable space. Naturally, such exploits do not work in the real world. Our experiments show that it is possible to tune simulation parameters to improve sim2real predictivity (e.g. improving $SRCC_{Succ}$ from 0.18 to 0.844), increasing confidence that in-simulation comparisons will translate to deployed systems in reality.

cs.CV↗