Skip to content

README.md

Latest commit

 

History

History
40 lines (27 loc) · 1.41 KB

README.md

File metadata and controls

40 lines (27 loc) · 1.41 KB

Master-Sim

Master-Sim is a project dedicated to developing high-fidelity synthetic data and pre-trained policies for precision assembly tasks in industrial robotics.

Overview

This project aims to bridge the gap between simulation and reality (Sim-to-Real) for tasks requiring delicate force control, such as peg-in-hole assembly.

Getting Started

Prerequisites

  • Python 3.8+
  • MuJoCo

Installation

pip install -r requirements.txt

Roadmap

See docs/DEVELOPMENT_PLAN.md for the detailed business and technical roadmap.

Quick Start (recommended)

Train (recommended: action labels as deltas):

python scripts/generate_expert_data.py --num-trajectories 200 --steps 200 --output data/expert_traj_delta.h5 --action-format delta
python scripts/train_bc.py --data data/expert_traj_delta.h5 --output models/expert_trained_delta.npz --epochs 100 --action-format delta

Demo:

python scripts/run_demo.py --model models/expert_trained_delta.npz --normalizers models/expert_trained_delta_normalizers.npz --num-episodes 20

Notes:

  • We recommend using action_format=delta (predicting joint deltas) for Behavior Cloning with IK-derived expert data: it produced stable policies with 100% success in our experiments (see docs/experiments/delta_vs_qpos_report.md).
  • To retain backward compatibility you can still generate or train using --action-format qpos.