GR00T-Pi0.5

MHBench decentralized baselines -- GR00T N1.7 and pi0.5

The two shipped MHBench decentralized policies, each one set of weights over every task and both robot roles. Only what evaluation loads is here: GR00T's merged standalone model dirs, and pi0.5's params/ and assets/. The optimiser state that would let you resume training is not (it is ~2.4x the size and eval never opens it).

• GR00T_N17/checkpoints/mhbench-multitask-unitree_g1x2_decentralized-joint-0/merged-<step>
• Pi_05/checkpoints/mhbench-multitask-unitree_g1x2_decentralized-joint-0/<step>/

Layout

Paths mirror a checkout, so a download drops straight into place:

GR00T_N17/checkpoints/mhbench-multitask-unitree_g1x2_decentralized-joint-0/
    merged-10000/ merged-20000/ merged-30000/ merged-40000/
Pi_05/checkpoints/mhbench-multitask-unitree_g1x2_decentralized-joint-0/
    10000/ 20000/ 30000/ 39999/     # each: params/ assets/

python baselines/scripts/mhbench_release.py download checkpoints --namespace Jebeom

lands them under baselines/XPolicyLab/policy/. GR00T additionally needs nvidia/Cosmos-Reason2-2B (~4.6 GB), which transformers fetches on first load -- the merged weights carry the backbone but not its tokenizer or processor.

Evaluating

EVAL_CKPT_NUM=40000 baselines/scripts/eval_launch.sh GR00T_N17 <task>

The GR00T weights are published under the untagged run name, which is the one the checkpoint resolver builds by default -- no CKPT_TAG needed. Upstream they were trained as mhbench-multitask-unitree_g1x2_decentralized-joint-0-pi05match, and that is the name they carry in wandb and in this benchmark's eval_results/ (GR00T_N17-decentralized-pi05match-seed0-ckpt<step>).

Success rate over 50 episodes per cell, from this benchmark's own eval_results/:

Training

GR00T is a LoRA finetune of nvidia/GR00T-N1.7-3B (alpha 32, dropout 0, adapters on the LLM's attention and MLP projections, the visual backbone and the action head), merged into standalone dirs by baselines/scripts/merge_lora_checkpoint.py. pi0.5 is an openpi finetune of pi05_base under the pi05_mhbench_multitask_decentralized TrainConfig. Both train on the flattened all-task dataset, where one row is one robot: 43 joint angles, one ego camera, and that agent's own instruction.

License

These weights are derivatives of nvidia/GR00T-N1.7-3B and nvidia/Cosmos-Reason2-2B (GR00T, merged) and of openpi's pi05_base (pi0.5), and carry those models' terms. Check them before redistributing.