τ₀-VLA LIBERO

The 60,000-step LIBERO checkpoint fine-tunes the τ₀-VLA pretrained low-level policy for end-effector control in the LIBERO benchmark.

Model interface

• Inputs: task instruction, front and wrist RGB images (224×224), and an 8D state containing end-effector xyz, axis-angle rotation, and two finger positions.
• Outputs: 10 actions, each with 7 values: delta xyz, delta axis-angle, and gripper.
• The model uses a unified 40D state/action representation. Active indices are 0:9 (xyz and 6D rotation) and 18 (gripper).
• Evaluation executes 8 actions before replanning. Long corresponds to libero_10.

Evaluation

Success rates (%) over 50 rollouts per task:

Download and run

Follow the LIBERO guide for the model and simulator environments. Check out the LIBERO implementation from PR #4:

git clone https://github.com/sii-research/tau-0-vla.git
cd tau-0-vla
git fetch origin pull/4/head
git switch --detach FETCH_HEAD
bash scripts/setup.sh
pip install -e '.[serve]'

hf download sii-research/tau-0-vla-libero \
  --local-dir checkpoints/tau-0-vla-libero

python -m deploy.libero_server \
  --model checkpoints/tau-0-vla-libero \
  --host 127.0.0.1 --port 8000 \
  --seed 7 --infer-mode eager --warmup-steps 1

In the separate LIBERO simulation environment, run all four suites:

for suite in libero_spatial libero_object libero_goal libero_10; do
  python -m deploy.libero.main \
    --args.host 127.0.0.1 --args.port 8000 \
    --args.task-suite-name "$suite" \
    --args.seed 7 --args.replan-steps 8 \
    --args.episode-start 0 --args.num-trials-per-task 50 \
    --args.video-out-path "outputs/libero_eval/$suite" || exit 1
done

Checkpoint files

The repository root is a complete inference export:

model.safetensors
config.json
run_spec.json
policy_manifest.json
processor_config.json
tokenizer.json
tokenizer_config.json
chat_template.jinja
finch_data_spec/libero-eef-robot-prompt-ft/
    spec.json
    components.json
    field_descriptions.json
    norm_stats.json

SHA256SUMS records checksums for the checkpoint files. The weight SHA-256 is e03870720cbddbd0f3be44ee929a5d23efb9bf9532f0ac1c1bf676224aacc8ec.

License

Code and model weights are released under the Apache License 2.0.