Egocentric Tactile and Proximity Sensors as Observation Priors for Humanoid Collision Avoidance

1. Problem Statement

Humanoid robots operating in dynamic, human-centric environments (e.g., crowded corridors, sports) require preemptive collision avoidance to ensure safety. While external cameras are common, they suffer from occlusion and high observation complexity. Consequently, egocentric sensors (tactile and proximity sensors distributed on the robot's body) are preferred due to their resistance to occlusion and low observation complexity.

However, a critical gap remains in understanding how to optimally shape the properties of these sensors (coverage geometry, signal type, and range) to serve as effective "observation priors." It is unclear which sensor configurations enable the most efficient learning of reactive, collision-free behaviors using Reinforcement Learning (RL), particularly regarding the trade-off between explicit object localization and raw proximity signals.

2. Methodology

A. Experimental Setup

• Robot Platform: The authors simulated an H1-2 humanoid robot equipped with 64 distributed sensors on its upper body.
• Task: An idealized dodgeball task where the robot must avoid incoming balls (radius 15 cm) thrown at velocities of 4–8 m/s from 2–6 meters away. Success is defined as maintaining an upright, contact-free posture for 3 seconds.
• Learning Framework:
  • Algorithm: Proximal Policy Optimization (PPO).
  • Architecture: A fixed, low-capacity two-layer feedforward network (32 units/layer) for both actor and critic. This design ensures performance differences are attributed to sensor morphology (the observation prior) rather than model expressivity.
  • Asymmetric Training: The critic has access to privileged linear base velocity, while the actor relies solely on egocentric sensor data and proprioception.
  • Control: The policy outputs target joint positions, converted to motor torques via a low-level PD controller to ensure compliant, stable motion.

B. Sensor Configurations (Ablation Study)

The authors systematically ablated sensor properties to test their impact on learning:

1. Coverage Geometry:
   • Field Sensors: Non-directional, spherical fields inspired by capacitive/acoustic sensors. They return a signal if any object is within range.
   • Ray Sensors: Directional beams inspired by Time-of-Flight (ToF) sensors. Each sensor consists of an 8x8 grid of beams (63° diagonal FoV).
2. Signal Types:
   • Perfect Localization: Returns the exact Cartesian coordinates of the ball (serves as an upper bound).
   • Relative Proximity: Returns the normalized distance to the closest point of the object within range.
   • Binary Detection: Returns a boolean (True/False) if an object is within range.
3. Ranges: Tested at 0.2m, 0.5m, 1.0m, and 2.0m.
4. Tactile Sensing: Binary contact sensors on all links were used for reward shaping. If any tactile sensor is activated, the episode terminates immediately, penalizing the agent for actual collisions.

3. Key Contributions

1. RL Framework for Whole-Body Avoidance: Developed a robust RL framework utilizing egocentric sensors to guide whole-body collision avoidance on a humanoid robot without relying on external cameras.
2. Sensor Property Characterization: Systematically characterized how sensor coverage, signal type, and range act as inductive biases that shape learned avoidance behaviors.
3. Signal Shaping Insights: Demonstrated that raw proximity measurements can effectively substitute for explicit object localization and that sparse, non-directional signals often outperform dense, directional ones in sample efficiency.

4. Key Results

A. Sensor Performance vs. Localization

• Perfect Localization: Naturally yielded the highest performance across all ranges but requires complex state estimation (e.g., Kalman filtering).
• Relative Proximity: At long sensing ranges (>1m), relative proximity measurements performed nearly as well as perfect localization for both ray and field sensors. This suggests that knowing how close an object is is sufficient for reactive control if the range is large enough.
• Binary Detection: At short sensing ranges (<0.5m), binary detection signals outperformed relative distance measurements. This implies that coarse spatial awareness (knowing an object is "near") is sufficient for reactive dodging when the object is close.

B. Signal Density and Efficiency

• Sparse vs. Dense: Sparse non-directional signals (Field sensors) significantly outperformed dense directional signals (Ray sensors) in terms of sample efficiency.
• The "Ray Sensor" Bottleneck: Standard ray sensors (returning a full 8x8 depth image) performed poorly compared to field sensors. However, when the ray sensor output was reduced to a single minimum distance value (mimicking the dimensionality of a field sensor), performance improved drastically, matching the perfect localization baseline at 1m range.
• Conclusion: High-dimensional, dense signals (like full depth images) may hinder the agent's search through the solution space. Simpler, lower-bandwidth signals are more learnable.

C. Training Efficiency

• Configurations using field-based proximity sensors (ranges >1m) converged to stable policies in under 1,000 epochs (~20 minutes on an RTX 4090 GPU).
• In contrast, ray-based variants often failed to converge within the 3,000-epoch limit, highlighting the importance of signal shaping for practical deployment.

5. Significance and Implications

• Practical Deployment: The findings suggest that expensive, high-bandwidth state estimation (localization) may be unnecessary for reactive collision avoidance. Raw proximity signals are a cheaper, lower-latency alternative that yields comparable performance.
• Sensor Design: For real-world humanoid skins, designers should prioritize sparse, non-directional sensing over dense, directional arrays to improve sample efficiency and reduce computational load.
• Reward Shaping: The use of tactile contact detection as a hard termination condition provides a natural, unsupervised mechanism for reward shaping, effectively teaching the robot to avoid collisions without explicit "distance-to-object" rewards.
• Future Work: The authors note that future research should address sensor noise profiles (e.g., electrostatic interference in capacitive sensors) and validate these findings on physical hardware with real-world artificial skins.

In summary, the paper argues that simpler sensor representations (sparse, non-directional proximity) are not only sufficient for learning complex whole-body avoidance but are superior in terms of learning speed and data efficiency compared to complex, high-fidelity localization signals.