MINT-SJTU/Evo1_RoboTwin2_clean

Evo-1 · RoboTwin Checkpoint

Evo-1 policy trained on RoboTwin 2.0 (50 bimanual manipulation tasks, aloha-agilex embodiment, absolute 14-D joint control). A single multi-task policy trained on clean data only (50 tasks × 50 demo_clean demonstrations).

Overall success rate: 65.7% on RoboTwin-50 demo_clean (100 rollouts per task, horizon = 37). Per-task results: RoboTwin_evaluation/README.md.

Files

File	Description
config.json	Evo-1 model config (vlm_name = OpenGVLab/InternVL3-1B)
norm_stats.json	Per-task normalization stats — 50 keys robotwin_<task> under aloha_joint
mp_rank_00_model_states.pt	Model weights (~1.5 GB)

Usage

Full instructions: Evo-1 repo → 🧪 RoboTwin Benchmark.

# 1. point the Evo-1 server at this checkpoint dir (Evo1_server.py, __main__ block: ckpt_dir)
# 2. start the server
cd Evo_1 && PYTHONPATH=. python scripts/Evo1_server.py
# 3. copy the Evo-1 policy plugin into a RoboTwin checkout and run one task
bash eval.sh place_burger_fries demo_clean step_20000 0 0 ws://0.0.0.0:9000 37

arm_key / dataset_key are sent per-request by the RoboTwin client (aloha_joint + robotwin_<task>); the server reads them from the payload, so no server edit is needed for these.

⚠️ Evaluation recipe (all three matter)

1. horizon = 37
2. num_inference_timesteps = 50
3. Gaussian action smoothing, kernel = 9