robodojo_pi05

smoke_config

π₀.₅ (RoboDojo leaderboard checkpoint)

Serves the officially released RoboDojo π₀.₅ simulation checkpoints (leaderboard rank 4; paper arXiv:2607.04434 Table 1: average score 11.41 / success rate 6.91%, mean over 3 training seeds).

Checkpoint

All three training seeds are released in the RoboDojo-Benchmark/RoboDojo Hugging Face dataset:

# sparse-pull one seed (~7 GiB)
GIT_LFS_SKIP_SMUDGE=1 git clone --depth 1 --sparse \
  https://huggingface.co/datasets/RoboDojo-Benchmark/RoboDojo robodojo-ckpt
cd robodojo-ckpt
git sparse-checkout set ckpt/RoboDojo/Pi_05/RoboDojo-sim-arx_x5-joint-0
git lfs pull
export ROBODOJO_PI05_CKPT=$(pwd)/ckpt/RoboDojo/Pi_05/RoboDojo-sim-arx_x5-joint-0/59999

For seeds 1/2, also set config_name: pi05_base_aloha_full_sim_arx-x5_seed_<n> (the fork's per-seed train configs are identical except the seed, but keeping them matched preserves the exact upstream deploy path).

Run

vla-eval serve -c configs/model_servers/robodojo_pi05/pi05.yaml
# One smoke task:
vla-eval run  -c configs/benchmarks/robodojo/smoke_test.yaml
# Full protocol (one process per task — Isaac's SimulationContext is process-global,
# so a single `vla-eval run` on eval.yaml would only complete the first task):
scripts/run_robodojo_protocol.sh

Input contract mirrors XPolicyLab policy/Pi_05/model.py: three CHW uint8 cameras (cam_high ← RoboDojo cam_head, cam_left_wrist, cam_right_wrist), 14-D packed joint state, instruction prompt; the full action chunk executes open-loop before the next inference, as in upstream eval_one_episode.