Safe Human-to-Humanoid Motion Imitation Using Control Barrier Functions
Vision-based humanoid imitation with CBF-QP safety filter from a single camera.
Safe Human-to-Humanoid Motion Imitation Using Control Barrier Functions
Abstract
The authors present a vision-based framework enabling a humanoid robot to imitate human movements while actively avoiding collisions. Human skeletal keypoints are extracted from a single camera and converted to joint angles for motion retargeting. A Control Barrier Function (CBF) layer, formulated as a Quadratic Program (QP), filters imitation commands to prevent both self-collisions and human-robot collisions. The core claim is that CBFs provide a mathematically principled, real-time safety guarantee over raw imitation output.
Key Contributions
- Vision-only pipeline — skeletal keypoint detection from a single camera feeds motion retargeting (no motion capture suit or multi-view rig required)
- CBF-QP safety layer for real-time collision avoidance, filtering imitation commands before actuation
- Unified handling of self-collisions AND human-robot collisions in a single optimization formulation
Methodology
Three-stage pipeline:
- Keypoint detection — single camera → human skeleton via off-the-shelf pose estimator
- Motion retargeting — skeletal keypoints mapped to humanoid joint angles
- CBF-QP safety filter — raw joint commands passed through a QP that enforces barrier constraints on proximity to obstacles (both the robot's own links and the human operator), outputting the closest feasible command to the original imitation target
The CBF framework provides forward-invariance guarantees: if the state starts in the safe set, CBFs mathematically guarantee it never leaves, subject to feasibility of the QP.
Results
Simulation results validate the effectiveness of the proposed framework for real-time collision-aware motion imitation. Specific quantitative benchmarks are not provided in the abstract.
Limitations
- Validation is simulation-only in the abstract — no hardware demonstration described
- Single camera implies depth ambiguity risks in keypoint estimation
- CBF-QP feasibility not discussed for aggressive motions where the barrier set becomes very tight
Subject Classification
- cs.RO (Robotics)
- eess.SY (Systems and Control)
Source: Safe Human-to-Humanoid Motion Imitation Using Control Barrier Functions — Cai, Abanes, Evangeliou, Tzes, April 2026