ALERT: Extending Humanoid Motion Trackers with a General and Lightweight Human Avoidance Pipeline

TL;DR Adds gated low-rank branches to a frozen, pretrained humanoid motion tracker and trains them with LiDAR-based repulsive potentials, so limbs near people bend away while the rest of the motion stays on track. On a Unitree G1 it completed all 20 trials of each of four dances, kicks and teleoperation runs without collision or fall.
What they do
Humanoid motion trackers can dance, kick and follow a teleoperator, but they do not perceive the people around them. ALERT adds human avoidance to an existing pretrained tracker without retraining it. Only limbs at risk deviate, and tracking of the reference resumes once the person is clear.
How it works
- Perception. An omnidirectional LiDAR detects people. A particle filter tracks each person’s center, and a convex envelope is scaled online to cover their point cloud.
- Potential and flow. Each envelope defines a smooth repulsive potential, summed over people. At 21 points on the robot’s body, the potential’s descent direction is added to the reference motion’s velocity to form a target velocity flow.
- Adaptation. The backbone is frozen. Low-rank (LoRA) branches gated by the potential features are trained with PPO on 500 motion sequences, using a reward for matching the target flow plus a collision penalty. The same recipe is applied to SONIC, TWIST2 and BFM-Zero.
Results
- Unitree G1 running adapted SONIC: 100% success over 20 trials each of Ninja Dance, Disco Dance, Kick and teleoperation, with at least 3 cm clearance from reconstructed human surfaces.
- Simulation on SONIC, 100 trials per task: 100% success versus 95% for CSF, 83% for ConstrainedMimic and 0% for plain SONIC, with lower tracking error on the avoiding limbs (84 mm versus 142 and 172 mm).
- Without people nearby, the adapted trackers stay close to the originals (SONIC global body-position error 52.3 to 54.7 mm).
- Ablations: full fine-tuning instead of LoRA drops success to 51%, center-distance inputs instead of potentials to 60%, and removing the flow reward to 79%.
Limitations
Convex envelopes cannot represent gaps between a person’s limbs, so close-range avoidance can be conservative. Tracking errors are higher on hardware than in simulation, and only the SONIC backbone was tested on the real robot.
Figures




Abstract
Whole-body motion tracking has enabled humanoid dancing, athletic motions, and teleoperation, providing a foundation for capability demonstrations and manipulation data collection. However, existing tracking policies lack awareness of nearby humans and proactive avoidance capabilities, while motion substitution and local constraint-based corrections may compromise reference motion integrity and tracking continuity. We propose ALERT, a general lightweight pipeline for extending motion tracking to shared human-robot environments. Through lightweight adaptation, it integrates human perception and proactive avoidance into pretrained tracking policies, balancing avoidance demands with reference motion tracking while preserving existing motion capabilities. Simulation and real-world experiments across tracking backbones demonstrate reliable collision-free execution through localized human avoidance, preserving whole-body coordination and enabling rapid recovery of reference tracking, while retaining the ability of pretrained policies to track unseen motions.
Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.
