Learning Social Robot Navigation By Sensing Human Legs

Technical Summary: Learning Social Robot Navigation By Sensing Human Legs

Problem Statement
Social navigation for mobile robots in populated environments faces a critical gap between sensor reality and algorithmic modeling. Most mobile robots utilize 2D LiDAR sensors mounted 10–20 cm above the ground. At this "ankle height," the sensor does not perceive pedestrians as unified cylindrical bodies or discs; instead, it detects two small, independently moving point clusters corresponding to the lower legs. Furthermore, standard LiDAR beams often miss the shoe protruding below the scan plane, creating a blind spot for obstacles extending forward from the ankle.

Despite this physical reality, the majority of learning-based navigation methods approximate pedestrians as simple circular agents, ignoring leg dynamics and foot geometry. This simplification creates a "sim-to-real" gap where policies trained on disc-based models fail to account for the specific occlusion and motion patterns observed by ankle-height sensors. Additionally, existing approaches often rely on explicit pedestrian tracking and trajectory prediction, which are computationally expensive and prone to failure in real-world deployments, or they fail to explicitly enforce social norms like yielding.

Methodology
The paper proposes CALF (Convolutional Attention for Leg Features), an end-to-end neural architecture designed to interpret raw LiDAR scans directly into safe navigation commands without explicit pedestrian detection or tracking.

1. Simulation Environment (LegNav):
   To train CALF, the authors developed LegNav, a custom lightweight 2D simulator implemented in JAX for high-performance GPU execution. LegNav integrates three key components:

   • Non-Slip Gait (NSG) Model: A novel pedestrian dynamics model that represents humans not as single discs, but as two alternating feet with rectangular footprints. The model explicitly simulates the "stance" and "swing" phases of walking, anchoring the stance foot while the swing foot interpolates toward a dynamic target. This reproduces the alternating leg clusters and the shoe protrusion invisible to the LiDAR.
   • LiDAR Ray Tracing: The simulator performs ray tracing against the individual leg discs and shoe rectangles, generating raw sensor data that mimics the specific "two-cluster" signature of ankle-height LiDAR.
   • Human Dynamics: Pedestrian motion is governed by the Headed Social Force Model (HSFM), which decouples body orientation from velocity to simulate realistic non-holonomic human movement.

2. CALF Architecture:
   CALF is a hybrid end-to-end network combining:

   • 1D-CNN Encoder: A weight-shared convolutional encoder processes a temporal stack of $N$ LiDAR frames to extract leg-related features.
   • Temporal Self-Attention: A multi-head attention module processes the sequence of feature tokens to implicitly infer leg motion and environmental dynamics across time, bypassing the need for explicit tracking.
   • MLP Fusion: The spatio-temporal LiDAR features are fused with the robot's kinematic state and goal vector, processed by a Multi-Layer Perceptron (MLP), and routed to Actor-Critic heads.

3. Training Framework:
   The policy is trained using Deep Reinforcement Learning (DRL) with three algorithms: PPO (on-policy), SAC (off-policy with entropy), and TQC (distributional off-policy).

   • Reward Shaping: The reward function balances navigation efficiency (progress, speed) with social compliance. Crucially, a novel Yielding Reward is introduced to explicitly incentivize the robot to stop when a pedestrian enters a frontal yield zone.
   • Curriculum Learning: Training begins with simple scenarios and progressively unlocks complex crowd dynamics (e.g., bottlenecks, intersections) based on the agent's success rate.
   • Ghost Mode: During training, pedestrians are unaware of the robot (ghost mode), forcing the policy to learn active dodging. During evaluation, ghost mode is disabled to simulate real interaction.

Key Contributions

1. CALF Architecture: A novel hybrid end-to-end network (1D-CNN + Attention + MLP) specifically designed to process alternating leg features from 2D LiDAR data for social navigation.
2. Non-Slip Gait (NSG) Model: A simulation model that accurately reproduces human gait dynamics and the specific sensor-level occlusion effects (shoe protrusion, alternating leg clusters) that disc-based models ignore.
3. LegNav Simulator: A JAX-based, high-throughput 2D simulator capable of running thousands of parallel environments, enabling the training of deployment-ready policies in under one hour on a single consumer GPU.
4. Explicit Social Compliance: The introduction of a yielding reward mechanism and the "Yielding Score" metric to explicitly enforce socially compliant behaviors (stopping/waiting) rather than relying on them to emerge naturally.

Experimental Results
The authors evaluated CALF against classical model-based planners (DWA, MPPI) and other end-to-end RL baselines (TAGD, VanillaE2E) in unseen test scenarios.

• Algorithm Comparison: Among the DRL algorithms, PPO achieved the best balance, reaching a 95.1% success rate with the lowest active collision rate (3.3%) among learning-based methods. It also produced the smoothest trajectories (lowest angular jerk) and maintained high space compliance. SAC and TQC converged faster but exhibited lower sample efficiency in final performance or higher collision rates.
• Ablation Study (Leg vs. Disc): Replacing the NSG leg model with a standard disc approximation (CALF$_{PPO}$(discs)) caused a dramatic performance drop: the active collision rate surged from 3.3% to 24.5%, and the success rate fell to 73.1%. The disc-based policy learned to be aggressive and failed to yield, confirming that modeling leg dynamics is critical for safety.
• Comparison with Baselines: CALF$_{PPO}$ outperformed classical planners (DWA, MPPI) in success rate and efficiency while maintaining comparable safety. Unlike DWA, which often "froze" (high yielding score but high timeout rate), CALF stopped when necessary and resumed motion efficiently. It outperformed other RL baselines (TAGD, VanillaE2E) in safety and motion smoothness.
• Zero-Shot Real-World Deployment: The best policy (CALF$_{PPO}$) was deployed on a TurtleBot 4 with an ankle-height LiDAR without any fine-tuning. The robot successfully navigated indoor environments with 5–6 human participants, demonstrating proactive avoidance and compliant yielding behavior.

Significance
The paper argues that the primary motivation for this work is overcoming the discrepancy between idealized simulation (disc-based pedestrians) and actual physical sensor geometry (alternating leg clusters). The results demonstrate that simulating articulated leg dynamics at the sensor level is a key factor in reducing the sim-to-real gap.

The authors claim that their approach achieves an optimal trade-off between task performance, safety, and social compliance, addressing a common dilemma where classical planners are overly conservative (stalling) and existing learning-based methods often sacrifice safety or smoothness for throughput. The successful zero-shot deployment validates that training on a high-fidelity, leg-aware simulator allows policies to generalize to real-world physical constraints without domain adaptation.