ACT_sim_mixer_operating_stage1 (checkpoint 003000)

ACT policy for the RoboSynChallenge simulation task mixer_operating. This repository root is a complete LeRobot pretrained_model directory. The layout is identical to the official RoboSynChallenge/ACT_sim_mixer_operating repository.

Files

File	sha256
config.json	fab118b518716c9fbe8e18a13fe21dba05dd6526a900590bdece1697de363eea
model.safetensors	e7421dae4b9944b996f194e6ef1afbe2ae27326ed50c73e3cf395ad9206db06d
train_config.json	b4cae62009a350605564e18ecabe1e6e69ba7f67e46c7137869c89f9d1ac227c

train_config.json is the unmodified training record. It contains absolute paths of the training machine. The loader does not read these paths.

Model

• Architecture: ACT (LeRobot 0.3.3), ResNet18 backbone, dim_model 512, VAE on.
• Initialization: official checkpoint RoboSynChallenge/ACT_sim_mixer_operating.
• Fine-tuning: full parameters, lr 1e-6 (backbone 1e-6), batch size 8, 3000 steps, seed 1000.
• Data: official cobotmagic_Sim_mixer_operating (LeRobot v2.1, 1000 episodes, 316000 frames). The data is not included here.

Input and output

• observation.state, observation.qvel, observation.qf: float32, shape [14] each.
• observation.images.cam_high, cam_left_wrist, cam_right_wrist: float32, shape [3, 480, 640].
• action: float32, shape [14] (two arms, 7 values each, joint space plus gripper).
• Normalization: MEAN_STD for VISUAL, STATE, ACTION. Statistics are stored in model.safetensors.
• chunk_size 50, n_action_steps 50, n_obs_steps 1, no temporal ensembling.

Procedure: download and evaluate

1. Download the repository root:

   hf download mair-robosyn/ACT_sim_mixer_operating_stage1 --repo-type model --local-dir checkpoints/ACT_sim_mixer_operating_stage1/
   

2. Run the official ACT evaluation from the RoboSynChallenge repository:

   bash policy/act/eval.sh mixer_operating random checkpoints/ACT_sim_mixer_operating_stage1 0 --pytorch_device cuda --headless True
   

Results (team-internal, not organizer results)

• mixer_operating / random, headless, max 500 steps, LeRobot 0.3.3, PyTorch 2.7.0+cu128, RTX 5090.
• 50 episodes (seeds 240926–240931): 46/50 = 92%, mean 267.38 steps.
• Official initialization checkpoint, same protocol, 20 episodes: 17/20 = 85%.

Warning

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

Note: no network needed at load time

config.json sets pretrained_backbone_weights to null. The loader does not download torchvision ResNet18 ImageNet weights. All backbone weights come from model.safetensors. The weights are unchanged from earlier revisions.

Earlier revisions 5dffabefb1dde4c36b29ec28f94c66171f1e0ce1 and a9403db63292bac6188518416221ad968f9c2a2d set ResNet18_Weights.IMAGENET1K_V1. With those revisions, the loader needs network access or a cached resnet18-f37072fd.pth in $TORCH_HOME/hub/checkpoints/. The action output is identical in both cases (verified, max abs diff 0.0).