MINT-SJTU/Evo1_RoboTwin2_datascale

Evo-1 · RoboTwin-550 Checkpoint (randomized training)

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 the public StarVLA/RoboTwin-Randomized dataset — 50 tasks × 550 demonstrations per task (50 demo_clean + 500 demo_randomized), 27,500 demonstrations in total.

Overall success rate (100 rollouts per task per setting, horizon = 37):

Setting	SR
demo_clean	85.0% (4251 / 5000)
demo_randomized	83.2% (4161 / 5000)

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, per-setting normalization stats — 100 keys robotwin_<task>_clean / robotwin_<task>_rand 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
#    (8th arg `auto` appends the _clean / _rand suffix matching the task config)
bash eval.sh place_burger_fries demo_randomized step_60000 0 0 ws://0.0.0.0:9000 37 auto

arm_key / dataset_key are sent per-request by the RoboTwin client. This checkpoint's norm_stats.json is keyed per task and per data setting, so run eval.sh with dataset_key_suffix = auto (or an explicit _clean / _rand) — a bare robotwin_<task> key will not be found.

⚠️ Evaluation recipe (all three matter)

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