// momentum

tracking since W47

// why it's here

  • Active: last commit today
  • Built almost entirely by one person
  • Health 59/100 with only 36 stars

// how to help

// readme

SimHum co-trains a bimanual manipulation policy on simulated robot trajectories and real-world human demonstrations, then fine-tunes it on a small real-robot dataset. Simulation supplies robot-valid actions, and human data supplies real-world observations. With only 80 real-robot episodes per task, SimHum reaches 62.5% success on held-out OOD scenes across four bimanual tabletop tasks, 53.7% higher than Real only in absolute success rate. Under matched data-collection time, it improves over the best single-source pre-training baseline by 35.0% in absolute success rate.


Installation

git clone https://github.com/kaipengfang/SimHum.git
cd SimHum
conda env create -f env.yml
conda activate dit

Data Preparation

Convert raw HDF5 episodes to replay buffer format:

Read the full README on GitHub โ†—