tlchen/RoboSynChallenge_GrabtheCupCrew

RoboSynChallenge GrabtheCupCrew Checkpoints

This repository contains the inference checkpoints for the ten RoboSynChallenge preliminary tasks. The corresponding evaluation code is in the GrabtheCupCrew RoboSynChallenge repository.

The submission policy routes six tasks to SmolVLA and four to ACT. Each task loads its checkpoint from <checkpoint_root>/<task_name> and uses a fixed action window configured by the policy.

Download the checkpoint bundle

Download the published checkpoint bundle with:

hf download tlchen/RoboSynChallenge_GrabtheCupCrew \
  --local-dir /path/to/checkpoint

Each task directory includes the model weights and the configuration and pre/postprocessing files required for inference. SmolVLA task directories also include tokenizer assets.

Evaluation

Use the official evaluator and policy entrypoint from the code repository. Follow its root README for environment setup, including the required RoboSynChallenge and EmbodiChain revisions. From the RoboSynChallenge repository root, run one task as follows (replace GPU 0 if needed):

xvfb-run -a bash policy/Hybrid_ACT_SmolVLA/eval.sh \
  drawer_open_place random /path/to/checkpoint 0 \
  --max_episodes 100

Run a one-episode smoke evaluation by setting --max_episodes 1. The evaluator interface is:

bash policy/Hybrid_ACT_SmolVLA/eval.sh TASK SETTING CHECKPOINT_ROOT GPU_ID [EVALUATOR_OPTIONS...]

To evaluate all ten tasks:

for task in \
  drawer_open_place items_handover water_pouring table_rearrangement \
  click_bell mixer_operating manipulate_pipette item_assembly \
  sample_loading handle_basket; do
  xvfb-run -a bash policy/Hybrid_ACT_SmolVLA/eval.sh \
    "$task" random /path/to/checkpoint 0 --max_episodes 100
done

Task	Policy	Action steps
drawer_open_place	SmolVLA	10
items_handover	ACT	50
water_pouring	SmolVLA	10
table_rearrangement	ACT	50
click_bell	SmolVLA	10
mixer_operating	SmolVLA	10
manipulate_pipette	ACT	25
item_assembly	ACT	25
sample_loading	SmolVLA	10
handle_basket	SmolVLA	10