ReLoco: Unified Humanoid Locomotion and Fall Recovery via Behavior-Conditioned Motion Priors
TL;DR One proprioceptive policy makes a 64.5 kg Tiangong 3.0 humanoid walk, run, get up after falls and resume its command. It is trained with adversarial motion priors conditioned on whether the robot is upright. On hardware it got up after all 34 kick-induced falls in which the policy stayed in control.
What they do
A humanoid that falls has to get back up and then carry on with what it was told to do. ReLoco trains one proprioceptive policy that stands, walks, runs, gets up from any fallen posture and resumes its velocity command, with no controller switching. It runs on the Tiangong 3.0, a 1.70 m, 64.5 kg humanoid with 23 joints under policy control. The policy goes from Isaac Lab training to the robot without on-robot fine-tuning.
How it works
- Posture routing. In training, torso height and uprightness label each state as upright or not. The label picks the expert motions (locomotion or get-up) and the task reward (velocity tracking or torso height). The actor never sees the label.
- Conditioned motion prior. An adversarial motion prior (AMP) discriminator scores 10-frame motion windows together with their labels. It has a shared feature network and one output head per behavior.
- Refined references. Human motions are retargeted to the robot, then replayed by a tracking policy in simulation, and the robot’s own rollouts become the expert data. Across 43 sequences this cuts ground penetration from 29.0% to 0.75%.
- Training aids. Half the episodes start from get-up frames and random velocity pushes cause falls. An upward force on the head, 400 N at first, is withdrawn once get-up success passes 0.8.
Results
- In the simulation ablation, ReLoco reaches 96.8% get-up success, 96.96% push recovery with the command resumed, and 0.127 m/s velocity RMSE. A single-head conditional discriminator gets 90.2%, 93.4% and 0.149 m/s. Unconditional AMP tracks at 0.860 m/s RMSE, and task-only RL never gets up.
- From 1000 randomized fallen states, ReLoco gets up 98.3% of the time, against 99.9% for HoST and 49.7% for a single-reference tracker. HoST is faster (1.57 vs 2.02 s) but uses 10.7 kJ of mechanical work to ReLoco’s 3.1 kJ.
- On hardware, the robot covered about 10 min of randomized-command outdoor locomotion without falling. Kicked while standing, it fell 34 times with the policy still in control and got up every time. Over 30 timed trials it took 2.64 s on average.
- Kicked while walking or running, it kept going, or got up and resumed, in all 20 evaluable trials.
Limitations
Kick and push forces on the robot were not measured, so the paper gives no maximum recoverable force and no quantitative comparison across terrains. Hardware protection cut in on 8 of 47 standing kicks and on 1 running kick. Each simulation variant was trained once. HoST still gets up faster.
Videos
Figures



Abstract
Humanoid robots must recover from falls and resume commanded motion to sustain locomotion. Learning these behaviors in one policy requires supervision that accommodates both upright movement and ground-contact recovery. We present ReLoco, a framework that uses behavior-conditioned adversarial motion priors to learn locomotion, get-up from arbitrary fallen postures, and command resumption within one proprioceptive policy. During training, robot posture selects the expert pool and conditions a shared discriminator with two behavior-specific outputs. Task rewards and shared regularization support command tracking and recovery, while temporary assistance facilitates exploration. A simulation-based refinement stage supplies robot-executed motion references for this supervision. Ablations examine conditional supervision, output separation, and reference quality under matched settings. At a common training iteration, the full policy achieves $96.80\%$ independent get-up success and $96.96\%$ push-recovery success, including command resumption. Its torso planar velocity root-mean-square error is $0.1266$~m/s, $14.7\%$ lower than the conditional single-output variant, with similar joint-angle discrepancy measured by dynamic time warping. Deployment on the 64.5-kg Tiangong~3.0 humanoid demonstrates walking, running, get-up, and locomotion resumption after falls using one proprioceptive actor, without controller switching or on-robot policy fine-tuning.
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.
