mair-robosyn/ACT_sim_item_assembly_official

ACT_sim_item_assembly

ACT policy for the RoboSynChallenge simulation task item_assembly. The repository root is a complete LeRobot pretrained_model directory. This is the organizer-released checkpoint, redistributed by team mair-robosyn for delivery.

Files

File	Size (bytes)	sha256
config.json	2037	8d255efa3335bb6b0c964ff1d0cbc7ff02183fdb9708320cfa38653557dc9258
model.safetensors	206563032	69462a12dbb0fe9bee654226ca714778ca5f83e420258456721626befb15bdac
train_config.json	5472	a70347cf9f1838073edbc29ddbf9af89808344ed9446824f5f39446b6dbb6dd3

config.json differs from the organizer source in one field only: pretrained_backbone_weights is null (offline load requirement; the value "ResNet18_Weights.IMAGENET1K_V1" is not needed because model.safetensors replaces all backbone weights). Source: RoboSynChallenge/ACT_sim_item_assembly revision 58955958519157176712a1eff319018d78177ddb (source config.json sha256 137c7dbfc8aa0aafb6557234c67563135d78cd4b01355b22229945e467dc2e68).

Model

• Architecture: ACT (LeRobot), ResNet18 vision backbone, transformer dim_model 512 / 8 heads / 4 encoder / 1 decoder layers, VAE latent 32
• Initialization: official released checkpoint RoboSynChallenge/ACT_sim_item_assembly @ 58955958519157176712a1eff319018d78177ddb
• Training: lr 1e-5 (optimizer_lr and optimizer_lr_backbone), batch size 32, 80,000 steps, seed 1000 (official run item_assembly_bs64_c50_4090_23; values from train_config.json. This checkpoint was trained by the organizers, not by our team.)
• Data: cobotmagic_Sim_item_assembly (official), LeRobot v2.1, 1000 episodes, 318000 frames, 25 fps. The data is not included.
• W&B run: project robosynchallenge, run_id pxvfp8vk (entity not recorded in train_config.json)
• Code: official release by the organizers (training code not part of this delivery)
• Environment: official training environment (job name item_assembly_bs64_c50_4090_23); our evaluation environment: Python 3.10.12, lerobot 0.3.3, torch 2.7.0+cu128, RTX 5090

Input and output

• State inputs: observation.state, observation.qvel, observation.qf [14]
• Image inputs: observation.images.cam_high, observation.images.cam_left_wrist, observation.images.cam_right_wrist [3, 480, 640]
• Action: action [14] absolute joint positions (CobotMagic)
• Normalization: STATE MEAN_STD, ACTION MEAN_STD, VISUAL MEAN_STD
• chunk_size 50, n_action_steps 50

Procedure: download and evaluate

1. hf download mair-robosyn/ACT_sim_item_assembly_official --repo-type model --revision f4f1f467b0f25d34fa612f0d682807c0f0e77c99 --local-dir checkpoints/ACT_sim_item_assembly_official (model files verified at this revision; later commits update this model card only)
2. bash policy/act/eval.sh item_assembly random checkpoints/ACT_sim_item_assembly_official 0 --pytorch_device cuda --headless True

Results (team-internal, not organizer results)

• Protocol: setting random, 30 episodes, seed 240926, max_steps 500, host robosyn5090, official policy/act/eval.sh
• Success rate: 19/30 = 63.3% (2026-10-11)
• Reference: same checkpoint, 100 episodes (issue #21 evaluation series): 66/100 = 66.0%

Status

• UPLOADED: yes, revision f4f1f467b0f25d34fa612f0d682807c0f0e77c99 (re-download sha256 match for all 3 files, 2026-10-11)
• LOAD VERIFIED: yes, same revision (2026-10-11). Minimal load + dummy batch: action (1, 14) torch.float32 finite True.
• EVAL VERIFIED (team-internal): yes, same revision, 5 episodes (seed 240926, max_steps 500), SR 4/5 = 80.0%, 2026-10-11. Official policy/act/eval.sh on the re-downloaded directory wrote evaluation_metrics.json.

Warning

These results come from our local runs. The organizers did not verify them.