Research map / Reinforcement learning / Visual and representation RL
HumanoidArena: Benchmarking Egocentric Hierarchical Whole-body Learning
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.
Introduces HumanoidArena, a simulation benchmark of 7 leg-critical HOI/HSI tasks for evaluating high-level egocentric policies that output a 40D whole-body action executed by general motion trackers (TWIST2, SONIC), with in-GMT, cross-GMT and perturbation protocols.
Problem
Existing humanoid benchmarks do not evaluate the policy-tracker interface in hierarchical whole-body control, nor center leg-critical interactions, nor separate perception, grounding, execution and cross-tracker transfer shifts.
Approach
A shared 40D intermediate whole-body action interface between a high-level egocentric visuomotor policy and interchangeable low-level GMTs, enabling measurement of cross-GMT transfer as a first-class benchmark axis, together with leg-critical task suites and perturbation splits.
Key results
- In-GMT under TWIST2: FM best HOI/HSI averages 36.11% and 58.75%; under SONIC: DP best at 52.22% HOI and 65.83% HSI.
- Cross-GMT transfer is fragile: average drop 39.9% for TWIST2→SONIC and 36.0% for SONIC→TWIST2.
- pi0.5 has lowest fall rate in in-GMT setting (3.33% under TWIST2, 5.24% under SONIC).
Surprising findings
- Smaller cross-GMT drop does not imply better transfer when matched-GMT performance is already low (retention near zero).
- SONIC→TWIST2 has much higher fall rate (7.56% average) than TWIST2→SONIC (1.36%), though TWIST2→SONIC has similar task drops.
Limitations
- Only 7 tasks and two GMT backends (TWIST2, SONIC) evaluated
- Simulation-first benchmark; no sim-to-real validation reported in main text
- Low absolute success rates on several tasks and cross-GMT retention near zero
Experimental setup
Datasets: HumanoidArena teleoperation demonstrations (1,400 episodes)
Benchmarks: HumanoidArena
Metrics: success rate average fall rate cross-GMT drop relative transfer retention
Baselines: ACT Diffusion Policy Flow Matching pi0.5
Builds on: ACT Diffusion Policy pi0.5 TWIST2 SONIC
Closest papers by meaning
- Hierarchical World Models as Visual Whole-Body Humanoid Controllers (ICLR 2025 · similarity 0.91)
- Adversarial Locomotion and Motion Imitation for Humanoid Policy Learning (NeurIPS 2025 · similarity 0.90)
- Humanoid Horizon: Extending Task Horizon in Whole-Body Loco-Manipulation via Parallel Training, Dynamic Starting, and Reward Gating (NeurIPS 2026 · similarity 0.90)
- 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 0.88)
- From Experts to a Generalist: Toward General Whole-Body Control for Humanoid Robots (NeurIPS 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.