Research map / Reinforcement learning / Visual and representation RL
Universal Humanoid Motion Representations for Physics-Based Control
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Builds PULSE, a universal physics-based humanoid motion latent space distilled online from an all-AMASS motion imitator (PHC+) via a conditional variational information bottleneck with a learned proprioception-conditioned prior, and uses it for hierarchical RL downstream tasks.
Problem
Existing physics-based humanoid motion latent spaces (ASE, CALM) are learned from small, specialized motion datasets and cover only narrow movement styles, limiting their use for free-form motion tracking and diverse generative tasks.
Approach
Train an imitator (PHC+) that tracks all of AMASS, then distill its skills into a 32-d variational latent space with a learnable proprioception-conditioned prior and a temporal-smoothness regularizer; downstream policies act in this latent space with a residual action around the prior mean.
Key results
- PULSE reaches 99.8% success on AMASS-Train* imitation after distillation from PHC+ (100%), versus 98.9% for PHC on training data.
- On VR controller tracking on AMASS-Test, PULSE reaches 93.4% success, versus 37.6% for ASE-30Hz and 10.1% for CALM-30Hz.
- On the 14-sequence real-world Quest dataset, PULSE tracks all 14/14 sequences, matching training from scratch.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add data cleaning and progressive training modifications to PHC | PHC motion imitator on AMASS · AMASS-Train | success rate: 98.9 % -> 100 % | Helped |
| Distill PHC+ into PULSE with a variational bottleneck | PHC+ motion imitator · AMASS-Test motion imitation | success rate: 99.2% -> 97.1% | Hurt |
| Use PULSE instead of training from scratch | VR controller tracking policy · AMASS-Test | success rate: 93.4% -> 93.4 % | No clear effect |
| Add consecutive latent-code regularization (ablation) | PULSE · AMASS-Test VR controller tracking | success rate: 60.8% -> 93.4% | Helped |
| Add a learnable prior instead of a zero-mean Gaussian prior (ablation) | PULSE · AMASS-Test VR controller tracking | success rate: 45.6% -> 93.4% | Helped |
| Use residual actions relative to the learned prior (ablation) | PULSE downstream policy · AMASS-Test VR controller tracking | success rate: 18.1% -> 93.4% | Helped |
| Combine RL training with online supervised distillation (ablation) | PULSE · AMASS-Test VR controller tracking | success rate: 93.4% -> 71.0% | Hurt |
| Add online distillation from PHC+ (ablation) | PULSE motion imitator · AMASS-Test motion imitation | success rate: 32.6% -> 97.1% | Helped |
| Add progressive hard-negative updates during primitive training (ablation) | PHC+ primitive · AMASS-Test | success rate: 92.0% -> 98.5% | Helped |
| Replace ReLU with SiLU (ablation) | PHC+ primitive with progressive training · AMASS-Test | success rate: 97.8% -> 98.5% | Helped |
| Use a variational information bottleneck (ablation) | PULSE motion representation · AMASS motion imitation | success rate: 100% | Hurt |
| Use a spherical latent space (ablation) | Distilled motion representation · AMASS motion imitation | success rate: 100% | No clear effect |
Surprising findings
- Training the latent space with RL objectives combined with distillation hurts performance, and RL-only training from scratch failed to converge to a good imitator within 1e10 samples.
- Using the learned prior as a residual action is essential; without it, VR tracking success drops to 18.1%.
Limitations
- Distilled policy does not reach 100% imitation success; the variational bottleneck is lossy.
- VR controller tracking lags behind training from scratch, and random generation can get stuck in fallen or standing states.
Experimental setup
Datasets: AMASS QuestSim real-world Quest 2 dataset
Benchmarks: AMASS-Train* AMASS-Test* real-world Quest 2 dataset (QuestSim) speed reach strike terrain traversal
Metrics: success rate global MPJPE root-relative MPJPE acceleration error velocity error normalized return
Baselines: ASE CALM training from scratch PHC HuMoR
Builds on: PHC ASE CALM HuMoR PPO
Closest papers by meaning
- Hierarchical World Models as Visual Whole-Body Humanoid Controllers (ICLR 2025 · similarity 0.89)
- SUGAR: A Scalable Human-Video-Driven Generalizable Humanoid Loco-Manipulation Learning Framework (NeurIPS 2026 · similarity 0.89)
- H-GAP: Humanoid Control with a Generalist Planner (ICLR 2024 · similarity 0.89)
- Adversarial Locomotion and Motion Imitation for Humanoid Policy Learning (NeurIPS 2025 · similarity 0.88)
- Humanoid Horizon: Extending Task Horizon in Whole-Body Loco-Manipulation via Parallel Training, Dynamic Starting, and Reward Gating (NeurIPS 2026 · similarity 0.88)
- Robots Pre-train Robots: Manipulation-Centric Robotic Representation from Large-Scale Robot Datasets (ICLR 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.