KODA: Kinematics--Observation Data Alignment for Humanoid Learning from Human Egocentric Demonstration

Key points

• Training humanoid whole-body vision-language-action (VLA) policies requires costly robot demonstrations with aligned robot-view observations and controller-compatible whole-body state-action labels.1
• Egocentric human demonstrations offer scalable supervision but have misaligned views and sparse motion cues.1
• KODA first generates a simulator-checked whole-body trajectory using a sparse tracker with sampling-based command optimization.1
• We evaluate KODA on four Unitree G1 loco-manipulation tasks using two whole-body controllers, SONIC and HEFT.1

Sources (1)

• [1]KODA: Kinematics--Observation Data Alignment for Humanoid Learning from Human Egocentric Demonstration
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 15, 03:08 AM

  We present KODA (Kinematics-Observation Data Alignment), which converts egocentric RGB demonstrations into whole-body VLA training data using a reusable canonical trajectory to guide robot-view synthesis and generate controller-specific whole-body state-action supervision.

  Training humanoid whole-body vision-language-action (VLA) policies requires costly robot demonstrations with aligned robot-view observations and controller-compatible whole-body state-action labels.

Extractive summary: sentences quoted from the sources.