PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Carson Kohlbrenner, Niraj Pudasaini, William Xie, Naren Sivagnanadasan, Nikolaus Correll, Alessandro Roncone

arXiv:2610.31891v1cs.RO

TL;DR

Collision demonstrations can be costly and damaging, and it is unclear which proximity-signal representations best support sample-efficient avoidance. The paper compares eight representations, including lossy transforms and contact forecasting, in constrained dodgeball learning. Wide-coverage, low-resolution sensors outperform high-resolution alternatives in sample efficiency, while anticipating future contact reduces it.

  • Problem

    Because collision demonstrations can be costly and damaging, it remains unclear which whole-body proximity-signal representations support sample-efficient avoidance.

    Source: §1 · p. 1

  • Method

    The paper compares eight proximity-signal representations, including lossy transforms, latent compression, forecasting, and privileged state, in constrained end-to-end dodgeball learning.

    Source: §1 · p. 2, §2.2 · p. 3, §4.1 · p. 6

  • Results

    Wide-coverage, low-resolution sensors consistently achieved higher sample efficiency than high-resolution alternatives, while future contact information consistently reduced sample efficiency.

    Source: §7 · p. 9

  • Takeaways & Limitations

    The findings suggest sample-efficient avoidance may benefit more from compact, action-oriented peripersonal threat representations than from reconstructing the external world.

    Source: §5 · p. 9

    The generic sensors are assumed to have zero phase offset, latency, and noise, excluding these factors as design choices for sample efficiency.

    Source: §2.1 · p. 2

Abstract

from arXiv · show

Data-driven robots must learn to avoid collisions with as few demonstrations as possible to avoid collecting large datasets of destructive collision examples. We explore whole-body proximity sensors as inductive biases for learning reflexive collision-dodging behavior and find that lossy sensors with low resolution were stronger biases than high-resolution coverage. We also found that, perhaps unintuitively, anticipating when and where contact happens on the body of a robot reduced its learning sample efficiency. We investigated eight candidate representations in a reinforcement learning environment, which included 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.

1 INTRODUCTION

Collision demonstrations can be costly and damaging, making sample-efficient avoidance important. The paper investigates how whole-body proximity signals should be represented and whether anticipating contact improves learning.

1 INTRODUCTION

  • Collision data can be damaging and costly to collect, motivating ways to learn safe, reflexive avoidance from few demonstrations.The introduction frames this constraint around robots deployed in unstructured environments.Source: §1 · p. 1
  • Proximity sensors sense from the robot’s surface and can address occlusion, but their sample-efficient properties, such as range and signal representation, remain unsettled.The paper asks how to shape a proximity sensor’s signal and whether anticipating contact helps.Source: §1 · pp. 1–2
  • The paper proposes a heuristic favoring wide, overlapping low-resolution coverage and low signal dimensionality, and tests lossy signals and contact anticipation with 64 sensors on a simulated H1-2.It also proposes that anticipating when and where contact occurs can create sample-efficiency deficits, evaluated with a proximity world model.Source: §1 · p. 2

2 PROXIMITY SIGNAL REPRESENTATION

The section defines a generic proximity sensor and frames representation choice as a question of which collision-relevant information a compute-constrained learner can exploit. It introduces transformations and forecasting as candidate representations for empirical comparison.

2.1 GENERIC PROXIMITY SENSOR

  • A generic sensor activates when an object intersects its coverage; coverage is divided into individually activating receptive fields whose grid dimensions define resolution.The sensor suite observation is x_t ∈ R^(H×W×S); sensors are assumed to have no phase offset, latency, or noise.Source: §2.1 · p. 2

2.2 SAMPLE-CONSTRAINED USABLE INFORMATION

  • The section asks which representation makes collision-relevant information easiest to exploit under learner-capacity and training-budget constraints.Predictive V-information is used as an interpretive lens because a deterministic transformation may be easier to exploit under a restricted compute budget, even without increasing Shannon mutual information.Source: §2.2 · pp. 2–3
  • Candidate representations include lossy sensor-level transforms, latent compression, future-observation and contact forecasting, and an oracle revealing privileged incoming-object state.The PWM forecasts future observations from the current sensor input; if contact is predicted, the forecast time is contact time, otherwise it is a fixed horizon.Source: §2.2 · p. 3, Table 1 · p. 3
  • Steps to a target success rate and normalized area under the success curve evaluate how readily each representation is exploited.Figure 2 illustrates sensor depth-image grids, ball points crossing receptive fields, predicted future observations, contact-terminated rollouts, and encoder-decoder validation samples.Source: §2.2 · p. 3, Figure 2 · p. 4

3 CONTACT ANTICIPATION

The paper extends contact anticipation to whole-body collision avoidance with proximity sensors. Its proximity world model predicts future sensor observations and uses them to anticipate contact.

3 CONTACT ANTICIPATION

  • Contact anticipation predicts a distribution over future binary contact states from past observations, a whole-body proximity-sensor setting not previously explored for collision avoidance.Earlier learned contact-prediction work was validated in manipulation settings.Source: §3 · p. 4

3.1 PROXIMITY WORLD MODEL

  • The proximity world model generates future latent sensor observations, then predicts which observations will lead to contact.It combines an encoder, dynamics model, decoder, and contact-prediction module.Source: §3.1 · pp. 4–5
  • The PWM was pretrained on 6,000 expert-demonstration trajectories to approximate an idealized contact predictor and indicate the influence of anticipation on sample efficiency.The authors assume contact-prediction inaccuracy would be minimized in deployment.Source: §3.1 · p. 5

4 EXPERIMENTS

The experiments compare proximity-sensor representations and future-contact prediction for humanoid collision avoidance under a shared, constrained training budget. Low-resolution field sensors substantially outperformed raw high-resolution rays, while adding future-contact predictions reduced sample efficiency.

4 EXPERIMENTS

  • The task used a 64-sensor humanoid avoiding randomly timed, collision-course balls, with training runs sharing a fixed iteration budget.Source: §4 · pp. 5–6, §4.1 · p. 6

4.1 SENSOR REPRESENTATION ABLATION

  • 0.522 NAUC at 2 m was achieved by raw field sensors, versus a maximum NAUC of 0.02 for raw ray sensors.Source: Figure 6 · p. 7, §4.1 · p. 7
  • The perfect-localization true-position oracle, which observes the incoming ball’s true relative position, provided an upper-bound representation for both field and ray sensors, achieving NAUC 0.584 at 2 m.Source: §4.1 · p. 7
  • At 1 m, fmin(xray) outperformed raw ray observations by 5400%, despite discarding within-sensor direction.Source: §4.1 · p. 7
  • Learned latent compression improved sample efficiency over raw rays but remained below fmin; binary field detection beat direct proximity at 0.5 m.Source: §4.1 · p. 8

4.2 CONTACT PREDICTION ABLATION

5 DISCUSSION

Signal dimensionality was the strongest predictor of sample efficiency, while anticipating contact reduced it in the close-by, single-impending-collision task. The field-sensor representation resembles peripersonal space: overlapping body-distributed activations encode enough direction and distance for avoidance.

5 DISCUSSION

  • Signal dimensionality was the biggest predictor of sample efficiency, and anticipating contact distracted the robot from task-relevant information in the close-by, single-impending-collision task.Even a proximity world model trained on thousands of expert demonstrations with privileged pretraining data did not beat a simple per-sensor minimum-distance transformation.Source: §5 · p. 9
  • Field-sensor activations resemble peripersonal-space receptive fields: individually low-detail signals overlap across the body to encode sufficient direction and distance for avoidance.The experiment held sensor range fixed, while peripersonal-space range is known to adapt to nearby-object speed; dynamically adapting range remains future work.Source: §5 · p. 9

6 RELATED WORKS

Collision-avoidance approaches range from manually tuned quadratic-programming controllers to learned policies, imitation, and reinforcement learning. The paper identifies reducing training steps as a goal for real-time adaptation without manual tuning, demonstrations, or privileged simulators.

6 RELATED WORKS

  • Quadratic programming uses predefined constraints for reactive collision avoidance, but depends on careful tuning of dynamics and object estimates to maintain safety.Learned policies imitate experts, while reinforcement learning trains directly from observations and rewards but needs accurate simulation and can require privileged information and many trial-and-error steps.Source: §6 · p. 9
  • Reducing training steps is framed as a goal for collision avoidance that could support real-time end-to-end adaptation without manual tuning, expert demonstrations, or privileged simulators.Source: §6 · p. 9

7 CONCLUSION

Across eight candidate representations, wide-coverage low-resolution sensors yielded higher sample efficiency than high-resolution alternatives, while future-contact information consistently lowered it. The authors suggest low-dimensional, overlapping sensing and current-observation policies for computationally constrained reflexive motion.

7 CONCLUSION

  • Wide-coverage low-resolution proximity sensors consistently yielded higher sample efficiency than high-resolution alternatives across the tested representations.The authors relate this pattern to V-information theory under sample and compute constraints.Source: §7 · p. 9
  • Providing future information about when and where collision would occur consistently reduced sample efficiency and did not overcome the cost of extra signal dimensionality.Source: §7 · p. 9
  • The authors suggest wide-overlap, low-resolution sensing that preserves relative distance and policies conditioned only on current observations for computationally constrained reflexive motion.Source: §7 · p. 9
  • Generative AI tools assisted with synthetic-dataset setup, pipeline implementation, mathematical verification, source discovery, figure editing, language refinement, and software editing.The authors state that these tools were not used for data analysis, hypothesis creation, or interpretation, and that they reviewed AI-assisted work.Source: AI USE STATEMENT · pp. 9–10

A.1 DATA CURATION

The PWM training data combined expert demonstrations from policies using different observation types and training stages, with held-out validation and evaluation trajectories. The 60 Hz data was windowed at a 10 Hz stride, and contact-padding frames helped capture sparse pre-contact moments.

A.1 DATA CURATION

  • The PWM was pretrained on 6,000 expert trajectories spanning raw observations, perfect incoming-ball localization, and minimum sensor distance.Half the data came from intermediate policies at 500–2,500 training steps and half from fully trained 3,000-step experts; 15% was reserved for validation, with a separate 1,000-trajectory hold-out set from two other experts.Source: §A.1 · p. 12
  • Windowing the 60 Hz dataset at a 10 Hz stride made the dynamics model predict in 100 ms intervals to avoid failure when high frame rates obscure local dynamics.Source: §A.1 · p. 12
  • Contact-padding frames let training windows capture sparse moments leading up to contact within the full C + R sized window.This addressed the PWM failure mode of contact labels being too sparse for effective gradient passes.Source: §A.1 · p. 13

A.2.1 OVERVIEW

The proximity world model encodes sensor frames, jointly predicts future latent observations, and estimates first contact. The implementation trains its modules in sequence and holds them fixed during policy learning.

A.2.1 OVERVIEW

  • A 3D variational autoencoder encodes depth frames, conditional flow matching jointly generates future latents, and a causal head predicts first contact.The model uses three context frames and predicts nine future frames.Source: §A.2.1 · p. 13

A.2.2 IMPLEMENTATION

  • Weighted binary cross-entropy gives ranges below 1.6 m tenfold reconstruction weight, while within-sensor edge penalties preserve range boundaries.The encoder uses a light KL regularizer alongside reconstruction losses.Source: §A.2.2 · p. 14
  • Contact labels use only the first positive simulator step as the event, with later steps masked and pre-contact steps treated as survival.The contact head is trained on cached rollouts from the frozen dynamics model.Source: §A.2.2 · pp. 14–15
  • 96.9% future-contact prediction accuracy and 83.8% non-contact accuracy were achieved on the validation dataset.Figure 8 reports contact prediction accuracy on the validation dataset.Source: §A.2.2 · p. 15, Figure 8 · p. 15

A.3 TRAINING REGIME

  • The encoder-decoder, dynamics model, and contact head are trained sequentially and frozen in turn; training the proximity world model required a large trajectory dataset.The authors therefore note that the model itself is not sample-efficient for adaptive end-to-end collision avoidance on real data.Source: §A.3 · p. 15

Related papers

Loading 2610.31891v1…