Reproduce “TrajBooster: Boosting Humanoid Whole-Body Manipulation via Trajectory-Centric Learning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jiacheng Liu, Pengxiang Ding, Qihang Zhou, Yuxuan Wu, Da Huang, Zimian Peng, Wei Xiao, Weinan Zhang and 3 more
TrajBooster introduces a framework that uses end-effector trajectories to transfer whole-body manipulation skills from large wheeled-humanoid datasets to bipedal humanoids, addressing data scarcity for training Vision-Language-Action (VLA) models. The approach enables a bipedal robot to perform complex tasks with higher success rates and improved generalization using significantly less real-world fine-tuning data.
Key insights
- End-effector trajectories serve as an effective, morphology-agnostic interface for abstracting and transferring complex whole-body manipulation skills across robots with diverse physical structures.
- A hierarchical whole-body control policy, specifically a composite arm-worker-manager structure, can efficiently translate high-level trajectory goals into robust low-level whole-body actions for bipedal humanoids.
- Post-pre-training VLA models with extensive retargeted simulation data significantly accelerates real-world adaptation, allowing for robust performance and enhanced generalization with substantially less robot-specific data collection.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.