YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
RoboSynChallenge ten-task checkpoints
Download access
This is a public gated model repository with manual approval. File names and the model card are public; downloading weights requires an approved Hugging Face user account. The HF user RoboSynChallenge is pre-approved for organizer evaluation. Organizers must authenticate as that user (hf auth login, legacy huggingface-cli login, or HF_TOKEN); other accounts require separate approval. Do not include tokens in code or submission forms.
Authenticated download (replace HF_REVISION with the fixed uploaded commit SHA listed in the code repository):
from huggingface_hub import snapshot_download
snapshot_download(repo_id="jagewo/robosyn-best-10tasks", revision="HF_REVISION", local_dir="./downloaded_release")
One selected checkpoint per task. All evaluation code and per-task configs are in the team's RoboSynChallenge GitHub repository under policy/g05/ and policy/pistar/. See that repository's root README for exact Hugging Face task links and installation instructions; the actual GitHub repository URL and fixed revision must be recorded before submission.
The task-to-weight mapping is checkpoints.json. Download the complete snapshot and run bash policy/g05/eval.sh TASK --release-root HF_DOWNLOAD_DIR for G05, or bash policy/pistar/eval.sh TASK --release-root HF_DOWNLOAD_DIR for PiSTAR. Supply --eval-python SIM_PYTHON and --model-python MODEL_PYTHON as needed. Each policy has deploy_policy.py, deploy_policy.yml, eval.sh and configs/TASK.yml.
G05 checkpoints require their .hydra/config.yaml, dataset_stats.json and action_tokenizer.pt alongside the selected .pt file. PiSTAR requires the complete params/ and assets/ directories; optimizer state, VLM value checkpoints and training datasets are not needed. Model licenses are included under runtime_assets/; original G05/Qwen/PiSTAR/base-model terms apply.
Use official RoboSynChallenge8818c82 with EmbodiChainv0.2.4.post1. CPU simulation/observation tensors and GPU inference are used. Video requires ffmpeg or imageio-ffmpeg. On the tested host, drawer and click_bell require setarch x86_64 -R; unaffected platforms can opt out with --enable-aslr. Evaluation outputs are saved before process exit. Local one-episode smoke checks verify runtime only; historical success rates predate simulator seed alignment and are not official held-out scores. Handover's historical unified-interface 0/100 is superseded by the current aligned evaluation, documented in docs/team_policy/HANDOVER_FIX.md in the code repository. Use positive advantage, beta=2.0, prediction/execution horizon 30 and CPU observations; the client verifies the live server beta. Weights are unchanged.
Current official-aligned local 100-episode validation: water_pouring 96/100, manipulate_pipette 64/100, table_rearrangement 90/100 and click_bell 83/100. Water and pipette use prediction32/execution10 and 10 flow-matching solver steps; pipette's 20-solver-step comparison scored60/100, so the task retains10. These are local results, not official held-out scores. See docs/team_policy/G05_REGRESSION_FIX.md and the archived metrics in the code repository for protocol differences and all comparisons. The G05 client requires its execution window to match the live server action_steps, validates explicitly configured solver steps, and maps a missing gripper prediction's observed fallback to physical units. The uploaded weights are unchanged; this release updates metadata only.
Training disclosure: official simulation demonstrations and task-specific synthetic rollouts with pretrained G05/pi05 models. Selected drawer round1_all uses all1,000 demonstrations and500 first-round rollouts, including successes and failures, with value-derived advantage labels. Second-round weights are not included. Review individual training records for exact task datasets before final submission.
Code repository: https://github.com/a1623194916/robosync. Use the final code commit recorded in the submission.
Handle_basket's complete comparison scored90/100 at prediction32/execution10 and79/100 at prediction32/execution32, both with10 solver steps. Keep execution10. All seven complete G05100-episode runs exited normally; complete results and tested per-task configs are recorded in checkpoints.json and the code repository report.