MINERVA — MINimal Efficient Robotic Vision-Action Policy

Public checkpoints for MINERVA, a compact, task-ID-conditioned policy for the standard 40-task LIBERO benchmark. The released 0.54M-parameter model uses a scratch CNN and an L1 action-chunk head. It does not use a language encoder, pretrained vision backbone, VLM, or iterative generative sampler at inference.

MINERVA studies a closed-set benchmark. It is not a general-purpose vision-language-action model and cannot execute unseen instructions.

Published result

Checkpoint	Parameters	Spatial	Object	Goal	Long	Average
t05_l1_0.54M	0.54M	96.8	99.6	97.4	89.2	95.75

Protocol: four suites, 10 tasks per suite, 50 episodes per task (2,000 rollouts), hard resets, evaluation seed 1000, temporal ensembling, and mujoco==3.3.2. The result is from one training seed. Aggregate tables and experiment notes are available in the source repository.

Evaluate the checkpoint

Use Python >=3.12,<3.14 and the locked environment from the source repository:

git clone --depth 1 https://github.com/k1000dai/MINERVA.git
cd MINERVA
uv python install 3.13
uv sync --python 3.13 --locked --extra libero

uv run hf download k1000dai/MINERVA \
  --revision 1b4fb1743f00a7d8eb87c7059c446447907d12bf \
  --include "t05_l1_0.54M/*" --local-dir ckpt

export MUJOCO_GL=egl
uv run lerobot-eval \
  --policy.path=ckpt/t05_l1_0.54M \
  --env.type=libero \
  --env.task=libero_spatial,libero_object,libero_goal,libero_10 \
  --policy.temporal_ensemble_coeff=0.01 --policy.n_action_steps=1 \
  --eval.batch_size=5 --eval.n_episodes=50 --env.max_parallel_tasks=1 \
  --seed=1000 --output_dir=eval_full

The MuJoCo pin is required: renderer changes in newer releases materially affect this scratch-CNN policy. One full evaluation needs approximately 15 GB of RAM.

The repository also contains the public t3C_2.75M teacher used to train the released model. See the training recipe and artifact manifest for pinned inputs.

Limitations

• The policy selects among 40 learned task IDs and does not generalize to unseen language tasks.
• The 95.75 result uses one training seed and one evaluation seed in one simulator.
• Performance is sensitive to renderer changes and visual or geometric perturbations.
• The checkpoint has not been validated on a physical robot.

License

The checkpoint weights and MINERVA source code are released under the Apache License 2.0. LIBERO, LeRobot, and the training dataset retain their own licenses and terms. MINERVA is a research fork of LeRobot, not an official LeRobot release.