Bridging Adaptivity and Safety: Learning Agile Collision-Free Locomotion Across Varied Physics

Yichao Zhong, Chong Zhang, Tairan He, Guanya Shi

Introduction

Deploying legged robots in dynamic real-world environments requires balancing three often-conflicting goals: agility, safety, and adaptivity. While a robot might be trained to run fast (agility) or avoid obstacles (safety), it often fails when physical conditions—such as floor friction or payload weight—change unexpectedly (adaptivity). For example, a high-speed policy might cause a robot to slip on a wet floor because it cannot perceive the change in friction. The Bridging Adaptivity and Safety (BAS) framework addresses this "trilemma" by integrating physical parameter estimation directly into a safety-aware control pipeline.

The BAS Framework

The BAS architecture builds upon the "Agile But Safe" (ABS) paradigm, introducing a layer of physical adaptivity. It consists of four integrated components:

  1. Agile Policy: A reinforcement learning (RL) policy optimized for high-speed goal-reaching.
  2. Recovery Policy: A robust policy designed to steer the robot away from imminent collisions or instability.
  3. Physical Parameter Estimator: This module analyzes proprioceptive history (e.g., joint velocities and angles) to explicitly estimate mass, friction, and Center-of-Mass (CoM) shifts.
  4. Reach-Avoid (RA) Value Network: Serving as a "safety guard," this network determines when to switch from the agile policy to the recovery policy. Crucially, the RA network is conditioned on the estimated physical parameters, allowing it to adjust safety margins dynamically (e.g., braking earlier when carrying a heavy load).

To ensure robustness, the authors employed Fusion Interpolation during training, mixing ground-truth physics with estimated values. Additionally, On-Policy Fine-Tuning was used to keep the estimator accurate even as the robot's behavior shifts between agile and recovery modes.

Experimental Results

The framework was validated using a Unitree Go1 quadruped robot in both simulation (IsaacGym) and real-world scenarios. Testing environments included slippery floors, rough carpets, slopes, and varying payloads up to 8kg.

Real-World Performance

In real-world "agility tests," BAS demonstrated a 19.8% increase in average speed compared to existing state-of-the-art methods. More importantly, it maintained high safety standards across diverse conditions.

BAS demonstrates its adaptivity by successfully navigating obstacles while carrying heavy payloads.

While non-adaptive baselines like ABS saw success rates drop to 34.38% in these challenging environments, BAS maintained an 81.25% success rate. The explicit estimation of physical parameters allowed the safety guard to trigger recovery maneuvers appropriately for the current dynamics, such as navigating slippery surfaces that would otherwise cause a crash.

The robot maintains stability and avoids collisions on slippery surfaces by adapting its safety margins to the environment.

Conclusion and Future Work

The BAS framework successfully bridges the gap between agility and safety by incorporating real-time physical adaptivity. By conditioning safety networks on estimated dynamics, robots can remain aggressive when safe and cautious when necessary. Future research will focus on integrating 3D vision for more complex obstacle detection and handling highly dynamic environments.

Summary updated

Abstract

Real-world legged locomotion systems often need to reconcile agility and safety for different scenarios. Moreover, the underlying dynamics are often unknown and time-variant (e.g., payload, friction). In this paper, we introduce BAS (Bridging Adaptivity and Safety), which builds upon the pipeline of prior work Agile But Safe (ABS)(He et al.) and is designed to provide adaptive safety even in dynamic environments with uncertainties. BAS involves an agile policy to avoid obstacles rapidly and a recovery policy to prevent collisions, a physical parameter estimator that is concurrently trained with agile policy, and a learned control-theoretic RA (reach-avoid) value network that governs the policy switch. Also, the agile policy and RA network are both conditioned on physical parameters to make them adaptive. To mitigate the distribution shift issue, we further introduce an on-policy fine-tuning phase for the estimator to enhance its robustness and accuracy. The simulation results show that BAS achieves 50% better safety than baselines in dynamic environments while maintaining a higher speed on average. In real-world experiments, BAS shows its capability in complex environments with unknown physics (e.g., slippery floors with unknown frictions, unknown payloads up to 8kg), while baselines lack adaptivity, leading to collisions or. degraded agility. As a result, BAS achieves a 19.8% increase in speed and gets a 2.36 times lower collision rate than ABS in the real world. Videos: https://adaptive-safe-locomotion.github.io.
BibTeX
@inproceedings{zhong2025bridging,
  author = {Yichao Zhong and Chong Zhang and Tairan He and Guanya Shi},
  title = {{Bridging Adaptivity and Safety: Learning Agile Collision-Free Locomotion Across Varied Physics}},
  booktitle = {L4DC},
  year = {2025},
  doi = {10.48550/arxiv.2501.04276},
}
Figure