EVO1 · LIBERO (Stage 2, 70k)

This is an EVO1 vision-language-action checkpoint trained on all four LIBERO suites with the two-stage LeRobot recipe: 5k steps of action-head-only training followed by joint VLM and action-head finetuning. This checkpoint was saved at step 70,000 of the 80k Stage-2 schedule and had the best overall result among the evaluated 10k, 40k, 60k, and 70k checkpoints.

• EVO1: https://github.com/MINT-SJTU/Evo-1
• Base VLM: OpenGVLab/InternVL3-1B-hf
• Dataset: lerobot/libero

Results

LeRobot LIBERO evaluation with seed 1000, 50 episodes per task, 10 tasks per suite, 14-action replanning, 448×448 observations, and 2,000 episodes overall:

These are single-seed results, not a multi-seed mean or confidence estimate.

Evaluation

Install LeRobot with the EVO1 and LIBERO dependencies:

pip install "lerobot[evo1,libero] @ git+https://github.com/huggingface/lerobot.git@e40b58a8dfa9e7b86918c374791599d070518d11"

The following evaluates task 0 of LIBERO Spatial. Run task IDs 0–9 to reproduce the suite result.

export MUJOCO_GL=egl
export PYOPENGL_PLATFORM=egl

lerobot-eval \
  --policy.path=zuoxingdong/evo1_libero \
  --policy.vlm_model_name=OpenGVLab/InternVL3-1B-hf \
  --policy.device=cuda \
  --policy.use_amp=true \
  --policy.vlm_dtype=bfloat16 \
  --policy.use_flash_attn=false \
  --policy.enable_gradient_checkpointing=false \
  --policy.n_action_steps=14 \
  --policy.num_inference_timesteps=32 \
  --policy.postprocess_action_dim=7 \
  --policy.binarize_gripper=true \
  --policy.gripper_threshold=0.0 \
  --policy.gripper_below_threshold_value=-1.0 \
  --policy.gripper_above_threshold_value=1.0 \
  --env.type=libero \
  --env.task=libero_spatial \
  --env.task_ids='[0]' \
  --env.camera_name=agentview_image,robot0_eye_in_hand_image \
  --env.camera_name_mapping='{agentview_image: image, robot0_eye_in_hand_image: image2}' \
  --env.control_mode=relative \
  --env.obs_type=pixels_agent_pos \
  --env.observation_width=448 \
  --env.observation_height=448 \
  --env.init_states=true \
  --env.episode_length=280 \
  --env.render_mode=rgb_array \
  --env.max_parallel_tasks=1 \
  --eval.n_episodes=50 \
  --eval.batch_size=1 \
  --eval.use_async_envs=false \
  --eval.recording=false \
  --seed=1000 \
  --output_dir=eval/evo1_libero/spatial-task-0

Use episode lengths 280 for libero_spatial and libero_object, 300 for libero_goal, and 520 for libero_10.

Training details

• Dataset: all four suites from lerobot/libero, revision a1aaacb7f6cd6ee5fb43120f673cebb0cfea7dd4, with the standard enabled LeRobot image transforms.
• Stage 1: randomly initialized action expert only, 5k steps.
• Stage 2: VLM and action expert finetuning, 80k-step cosine schedule; this release is step 70k.
• Both stages: global batch 128, learning rate 1e-5, warmup 1,000, weight decay 1e-3, dropout 0.2, gradient clipping 1.0, chunk size 50, image size 448, and padded state/action dimensions 24.
• Measured topology: two DDP processes with batch 64 per process on two H100 GPUs using bf16 AMP.
• Base VLM source revision: 014c0583a0d4bedf29fbe2dbff4f865eb998e171.
• Model weight SHA-256: c6ae0c7350d1810a064d16080137fdbc42d6dac5c6cf2cfb3874907efd959318.

Limitations

• Evaluation was run with one seed (1000).
• The reported behavior depends on 14-action replanning and the documented LIBERO gripper postprocessing.
• Loading the policy also downloads OpenGVLab/InternVL3-1B-hf unless it is already cached.

License

The EVO1 code is MIT-licensed. InternVL3 is released under MIT, while its Qwen language component is covered by the Qwen License. This model is therefore published with license: other and Qwen attribution. The lerobot/libero dataset is Apache-2.0.