Lastforever16/robosyn-ten-task-models

RoboSynChallenge: ten-task model weights and runtime assets

This repository provides the checkpoints and auxiliary runtime assets for all ten RoboSynChallenge tasks in one evaluation request. Task-specific checkpoints are used. The task-to-asset mapping and each task's run command are documented below.

• Weights: https://huggingface.co/Lastforever16/robosyn-ten-task-models
• Evaluation code: https://github.com/Mokk404/robosyn-submission
• Pinned, verified asset revision: da23c29b897034d0ca44f703d09985f21d629def

Checkpoint mapping for all ten tasks

Each URL below includes the fixed revision and the task subdirectory. These are complete task asset directories: hybrid policies need all listed auxiliary assets, not just one primary model file. The precise checkpoint paths and SHA-256 digests are recorded in ASSETS.json; the code repository's TASKS.json maps runtime arguments to those paths.

Task	Fixed-revision checkpoint / asset URL	Asset files	Selected policy and assets
table_rearrangement	table_rearrangement assets	2	ACT 080000 with physical gripper clipping
click_bell	click_bell assets	2	ACT 080000
water_pouring	water_pouring assets	2	ACT spatial-translation fine-tune 000200
handle_basket	handle_basket assets	2	ACT 080000
items_handover	items_handover assets	8	Dual ACT, RGB translation regressor, normalizers, gripper schedule and RBF prior
drawer_open_place	drawer_open_place assets	2	Placement-weighted ACT 002500
mixer_operating	mixer_operating assets	2	ACT retained baseline 100000
item_assembly	item_assembly assets	3	ACT 080000 and fixed bounded qpos residual
manipulate_pipette	manipulate_pipette assets	2	Pure ACT current1500
sample_loading	sample_loading assets	371	V89: four visual checkpoints, calibration and runtime trajectory assets
Total	All ten tasks	396

All tasks share this hosting repository, but that does not mean they share one model checkpoint. If a submission form requests one URL per task, use the corresponding row above. No duplicate upload is required merely because several tasks use the same repository.

Download and verify

Clone the code repository, then restore all assets into its root:

git clone https://github.com/Mokk404/robosyn-submission.git
cd robosyn-submission
python -m pip install huggingface_hub
python download_assets.py --repo-id Lastforever16/robosyn-ten-task-models --revision da23c29b897034d0ca44f703d09985f21d629def
python download_assets.py --verify-only

Keep every file at its original relative path. In particular, retain the complete items_handover auxiliary asset set and the sample_loading calibration and trajectory libraries. Do not flatten directories or move only the primary checkpoints.

The pinned revision is the asset snapshot verified before subsequent README-only updates. Those documentation updates do not require changing the asset revision or re-uploading model weights. main may be used intentionally for newer assets, but any new asset revision must be verified again.

Runtime and dependencies

Run on Linux with Python 3.11, an appropriate NVIDIA driver/CUDA-enabled PyTorch build, and the organizer-provided RoboSynChallenge / EmbodiChain simulator. This weights repository does not include the simulator or policy source code.

Recorded reference ACT versions include PyTorch 2.7.1, torchvision 0.22.1, LeRobot 0.3.3 and NumPy 1.26.4. Consult ENVIRONMENT.md, requirements-common.txt, task-specific requirements files, and each task's SUBMISSION.md in the code repository. The sample-loading controller has additional dependencies. Use ffmpeg when recording evaluation videos. Provision the official simulator before installing policy dependencies; the reference versions are not a newly certified universal runtime lock.

From the code repository root, set the actual runtime paths:

export ROBOSYNCHALLENGE_ROOT=/absolute/path/to/official/RoboSynChallenge
export PYTHON_BIN=/absolute/path/to/python
# Set this if the official runtime requires an explicit EmbodiChain checkout:
export EMBODICHAIN_ROOT=/absolute/path/to/EmbodiChain

Run command for each task

The commands below use the public random setting, GPU 0, 100 episodes and seed 0. The organizer may substitute held-out seeds, episode counts and settings; these example parameters are not restrictions on official evaluation.

bash eval.sh --task table_rearrangement --gpu 0 --episodes 100 --seed 0
bash eval.sh --task click_bell --gpu 0 --episodes 100 --seed 0
bash eval.sh --task water_pouring --gpu 0 --episodes 100 --seed 0
bash eval.sh --task handle_basket --gpu 0 --episodes 100 --seed 0
bash eval.sh --task items_handover --gpu 0 --episodes 100 --seed 0
bash eval.sh --task drawer_open_place --gpu 0 --episodes 100 --seed 0
bash eval.sh --task mixer_operating --gpu 0 --episodes 100 --seed 0
bash eval.sh --task item_assembly --gpu 0 --episodes 100 --seed 0
bash eval.sh --task manipulate_pipette --gpu 0 --episodes 100 --seed 0
bash eval.sh --task sample_loading --gpu 0 --episodes 100 --seed 0

For a preliminary path check, append --dry-run. For a one-episode simulator smoke test, change --episodes 100 to --episodes 1. Additional official evaluator arguments may be appended after --, for example:

bash eval.sh --task table_rearrangement --gpu 0 --episodes 1 --seed 0 -- --eval_video_log True

The unified entrypoint uses the packaged task policy and its existing action conversion. Table's physical gripper handling is task-specific and is not applied globally to other tasks. The entrypoint does not modify official task logic, randomizer code or success conditions.

Data sources and compatibility disclosures

Available training provenance records simulated manipulation demonstrations, including released RoboSynChallenge simulation data and task-specific additional simulated demonstrations. Hybrid components also use visual pose supervision and packaged trajectory/calibration assets. Dataset identifiers and available source records are documented in DATA_SOURCES.md and DATA_PROVENANCE.json in the code repository. Exact episode counts, revisions and data provenance are not fully established for every component; no unsupported counts or permissions are claimed here.

Some ACT policies require robot qvel/qf in addition to RGB and qpos. Availability of these fields must be checked against the organizer's actual runtime. The sample_loading controller uses env.unwrapped.robot FK/IK functionality; written organizer confirmation of this access remains outstanding. See the code repository's remaining-issues documentation.

Verification scope

On 2026-10-11, all 396 required runtime assets at revision da23c29b897034d0ca44f703d09985f21d629def were checked against the local asset manifest: all ten tasks were present, file sizes and SHA-256 digests matched, and the repository was publicly accessible without authentication. ASSETS.json and SHA256SUMS support independent integrity verification of their listed assets.

This confirms upload integrity. It does not certify a fresh GPU rollout of the unified launcher, organizer approval, or held-out evaluation performance. Historical evaluation results belong to the experiments documented in the code repository.