acceptodds
Withdrawn from ICLR 2027

Move, don't think: sample-efficient collision avoidance with lossy proximity sensors

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Robots must learn to avoid collisions with as few demonstrations as possible to overcome the challenge of collecting large datasets of potentially destructive collision examples. By using egocentric proximity sensor observations as inductive biases, we train a humanoid robot to learn collision avoidance in fewer steps using lossy, low-dimensional sensors with wide overlapping coverage rather than with high-resolution, spatially accurate sensing of the robot's surrounding space. We also find that, perhaps unintuitively, anticipating when and where contact happens on the body of a robot harms its learning sample efficiency. We investigate eight candidate representations for sensor observations as an inductive bias for learning collision avoidance using reinforcement learning, which include latent compression, future contact anticipation, lossy corruption, and privileged state estimation. The candidate representations were tested by training over 200 humanoid robots to play a simplified version of dodgeball end-to-end under an intentionally constrained sample and compute budget.

Primary area:
applications to robotics, autonomy, planning
Keywords:
contact anticipation, egocentric proximity sensing, collision avoidance, safe human-robot interaction, world models, sample efficiency, learned reflexes

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_6RiqOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies how proximity-sensor representations affect the sample efficiency of learning collision avoidance. Using a simulated humanoid dodgeball task, it compares raw observations, lossy transformations, learned compression, and future-contact prediction. Under the tested policy architecture and training budget, low-dimensional sensing and simple minimum-distance features support faster learning, while adding future-contact information generally slows it down.

Critical strengths
  1. The paper asks what information a resource-constrained policy can actually use, rather than assuming that more detailed sensing or future predictions will always help. This is a worthwhile question for robot learning.
  2. The experiments cover simple transformations, learned representations, and oracle observations. Multiple seeds and learning-curve metrics provide a more informative evaluation than final success rates alone.
Critical weaknesses
  1. Videos of successful and failed trials would help show how quickly the robot reacts and how it avoids the ball. The static examples leave me unsure how much the task actually requires detailed spatial sensing or anticipation, which matters for interpreting the results.
  2. The small two-layer MLP fits the paper’s setting, but it is unclear whether low-dimensional representations are better for avoidance learning or simply easier for this network to handle. Testing a few network sizes would help clarify how much the findings depend on this choice.
  3. Predictions are added as extra policy inputs, so slower learning could mean either that future information adds little value or that the policy struggles to use it. Comparing different ways of incorporating predictions would help separate these explanations; otherwise, the conclusions should be limited to the current setup.
Questions for authors
  1. For each representation, what proportion of failures results from ball collisions versus loss of balance? How does this breakdown change during training?
  2. How is each observation representation normalized before entering the policy? Are PPO hyperparameters shared across representations or tuned separately?
Reviewer_AHjiOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper considers the problem of training policies via RL for controlling humanoid robots for collision avoidance, focusing on using egocentric proximity sensors. The paper focuses on both understanding the sensor type (ray vs field) and representation of the sensor that can lead to better RL performance, as well as if predicting future sensor and contact information, and then conditioning the policy on this information, can improve performance. The paper then evaluates different sensor information by training small policies to control simulated unitree humanoid robot to avoid incoming balls, finding the lower-resolution sensors perform better. This setup is then augmented with a proximity world model (PWM), consisting of an variational autoencoder/decoder, latent dynamics trained with flow matching, and a separate head to predict contact, where each component is trained offline prior to online RL. The PWM uses the ray sensors, and provides the input to the policy for the latent representation of current observation and predicted future information. The paper finds that the PWM does not improve the base approach.

Critical strengths
  • The paper investigates a range of sensor configurations, and also proposes a (proximity) world model architecture that could help anticipate future sensor observations and contact, which seems like an interesting idea for avoiding collision with dynamic objects.
Critical weaknesses
  • A major weakness I find with this paper (relating to contribution 1,2,3) is the use of relatively small policies (2 hidden layers with 32 hidden units); this may confound the findings of which sensor observations are more effective where “signal dimensionality was the biggest predictor of sample efficiency”. With a small network size, it may only be able to handle the smallest types of sensors. This may be similarly why the PWM information is not useful, as this further increases the dimensionality of the input while maintaining the small network size. Therefore, I think it is important to also study larger networks (e.g. 512 hidden units).
  • Contribution 2 seems like an overgeneralization from the negative world modeling results: “anticipating when and where contact will happen can create sample-efficiency deficits for learning collision avoidance”. In many prior works, world models, such as in model-based RL can improve sample efficiency. This may be a product of a sub-optimal configuration, rather than anticipation/world-models not being useful in general.
  • I think some of the discussion in section 2.3, relating to V-information is a little unclear, and V-information is not measured later in the paper. If the purpose of this section is to motivate that a more simple, or low-dimensional representation helps a capacity constrained learning, it may be more clear to directly state this. If the paper would want to go further in the direction of understanding the relationship between constrained capacity and sensor information, the paper should perform a sweep over various capacity and sensor combinations. This could show that there may be some sensors that work across different learner capacities, some that scale, or others that are ineffective across all learner capacities.
  • The claim that the findings in the paper “Our findings are further supported by the neurosciences, where fast multisensory neurons encode a peripersonal space” seems to be a tenuous justification of the findings.
  • The scope of the experiments is relatively narrow. The paper considers a single simulated scenario, avoiding an incoming ball, while trying to maintain a relatively stationary position. The episodes also seem short, in that they end when one ball collides, and end if the robot survives for 3 seconds. It seems that longer horizon and more diverse tasks should be studied to inform decisions on observation sensor quality. The paper also does not study integration, or comparison to vision (RGB or RGB-D), and also does not study if their simulated findings transfer/inform design decisions for real sensors / robots.
Questions for authors

In Figures 5,9 why are there no sensor readings when contact occurs for the ground truth? Wouldn’t the ball still be able to seen by the sensors at this point? Is the dynamics model trained to predict this (no sensor readings) as a target? if so, it seems like this is a discontinuity that could harm the learning of the PWM.

Reviewer_2AD8OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates how different representations of whole-body proximity sensor observations affect the sample efficiency of learning collision avoidance policies for humanoid robots. The authors compare low-resolution, wide-coverage field sensors with high-resolution ray sensors and evaluate several observation transformations, including minimum-distance pooling, binary activation, latent compression, and future contact prediction using a proximity world model (PWM). The experiments are conducted in a simulated humanoid dodgeball task using PPO under a fixed training budget. The main findings are that compact and lossy sensor representations can substantially improve learning efficiency, while augmenting policy observations with predicted future contact information can slow down learning.

Critical strengths
  1. The paper addresses an interesting and practically relevant question, about how should sensor observations be represented to improve learning efficiency under limited computational resources.
  2. The experimental results demonstrate substantial improvements from simple observation transformations. In particular, per-sensor minimum-distance pooling significantly outperforms raw high-dimensional ray observations, which is a useful empirical finding.
  3. The paper studies multiple sensor configurations and representation choices under a relatively consistent RL framework. The inclusion of predictive world-model representations provides an interesting perspective on the usefulness of anticipation for reactive control.
Critical weaknesses
  1. Experimental setting not fully fair or controlled. The field and ray sensors differ not only in resolution but also in their sensing coverage. This FoV gap could lead to performance gap, which affects the main findings of this paper.
  2. The future action predicting module is not working very well, which may lead to bad PWM performance, and this cannot support the strong claim that future contact anticipation harms sample efficiency. Figure 8 reports 96.9% accuracy on positive contact labels and 83.8% on non-contact labels. Based on the reported counts, the contact prediction precision is approximately 46.1%, indicating a substantial number of false positive predictions. It is therefore difficult to determine whether the reduced learning efficiency arises from anticipation itself or from inaccurate and potentially misleading predictive features.
  3. The task setting and generality of the findings are limited. The evaluation is restricted to a simplified dodgeball simulation in which projectiles are launched toward the robot along predefined collision courses. In this setting, current proximity observations may already provide sufficient information for reactive avoidance, reducing the potential benefit of future prediction. It remains unclear whether the conclusions generalize to dynamic obstacles, multiple simultaneous threats, changing velocities, or other settings where anticipating future motion is essential.
Questions for authors
  1. Can the authors provide a controlled resolution ablation under identical sensing coverage, rather than comparing wide-FoV field sensors with narrow-FoV ray sensors?
  2. Does future contact anticipation still reduce learning efficiency when the policy receives ground-truth future observations or perfectly accurate contact information? How much of the degradation can be explained by increased feature dimensionality alone?
  3. Given the class-specific accuracy reported in Figure 8, what are the contact predictor's precision, first-contact timing error, false-alarm rate, and calibration? Could false positive predictions explain the degraded downstream performance?
  4. Do the main findings persist in scenarios involving multiple projectiles, varying object velocities, or non-collision trajectories?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.