Dexterous Safe Control for Humanoids in Cluttered Environments via Projected Safe Set Algorithm

TL;DR

This work tackles the challenge of dexterous safety for high-DOF humanoids in cluttered environments by promoting multi-constraint safety with limb-level geometry. It introduces the Projected Safe Set Algorithm (p-SSA), a tuning-free method that decouples feasibility from task optimization by projecting potentially infeasible safety constraints to a nearest feasible set and then solving a safe-control problem within that set. The paper also presents Relaxed Safe Set Algorithm (r-SSA) as a precursor, analyzes sources of infeasibility (inherent, method, and kinematics), and demonstrates through simulation and real hardware (Unitree G1) that p-SSA achieves near-optimal performance with minimal safety violations across diverse tasks. The approach provides practical safety guarantees in dense environments and shows strong potential for safe teleoperation and dexterous manipulation in real-world humanoid robotics.

Abstract

It is critical to ensure safety for humanoid robots in real-world applications without compromising performance. In this paper, we consider the problem of dexterous safety, featuring limb-level geometry constraints for avoiding both external and self-collisions in cluttered environments. Compared to safety with simplified bounding geometries in sprase environments, dexterous safety produces numerous constraints which often lead to infeasible constraint sets when solving for safe robot control. To address this issue, we propose Projected Safe Set Algorithm (p-SSA), an extension of classical safe control algorithms to multi-constraint cases. p-SSA relaxes conflicting constraints in a principled manner, minimizing safety violations to guarantee feasible robot control. We verify our approach in simulation and on a real Unitree G1 humanoid robot performing complex collision avoidance tasks. Results show that p-SSA enables the humanoid to operate robustly in challenging situations with minimal safety violations and directly generalizes to various tasks with zero parameter tuning.

Figures (12)

• Figure 1: Application of dexterous safe control for humanoids in cluttered environments. (a) A safe teleoperation task where the Unitree G1 humanoid mirrors Apple Vision Pro (AVP)-tracked human arm motions to manipulate objects inside a cabinet while avoiding both obstacle and self-collisions. (b)-(1) p-SSA blocks excessive squeezing near the cabinet (frames 2–3) and allows safe input in frame 4. Obstacles triggering p-SSA intervention are highlighted in blue in AVP view. (b)-(2) p-SSA prevents collisions when both arms are inside the cabinet. This poses a more complex problem due to multiple concurrent safety constraints within a confined space. (c) Simulated goal-reaching (green) with collision avoidance with multiple obstacles (gray) via p-SSA.
• Figure 2: Possible scenarios where \ref{['prob:multi_safe_control']} can be infeasible. The humanoid should avoid collision with all obstacles (e.g., planes and spheres) in gray.
• Figure 3: Unitree G1 humanoid robot in MuJoCo simulation performing safe wrist location tracking. The humanoid tracks the goal (green) with its wrist while preventing collisions between robot bodies (black) and obstacles (gray) and self-collision. There are three active control constraints (blue) triggered by collision bodies being too close, while two of them are infeasible and relaxed by p-SSA (purple). The infeasibility is caused by the right arm trying to avoid both the obstacle and the torso at the same time.
• Figure 4: Comparison of safe control methods in G1FixedBase_DO_v0 task. Spheres and lines follow the convention in \ref{['fig:system']}. When an obstacle moves near the left elbow (frame 2), the QP becomes infeasible. In that case, p-SSA (top) generates control to minimize violation to control constraints, resulting in less violation (purple connection) than r-SSA (middle). Naive SSA (bottom) does not handle infeasible control constraints (thick red connection), leading to collisions (red spheres).
• Figure 5: Performance comparison under G1WholeBody configuration.