Kite’s task-trained LeRobot ACT policy for red-cup picking on a bimanual WidowX AI station. This baseline-only checkpoint uses 17 simulated episodes and 90,000 training steps; evaluation is by the publisher in simulation only.
Training code for this version is unconfirmed. Public code resources do not establish access to the exact training implementation. View verification notes
Single cup task; author simulation only, no right-layout success or real-robot validation. Full training scripts, action units, control frequency, minimum VRAM and file-level license scope remain unverified. Available weights or code do not establish plug-and-play robot compatibility. Check the required configuration, data and interfaces in the source documentation.
Base model · Embodiment scope · Functional purpose · Record identity · Output · Pipeline task · Robot form · Task scope · Training data scale · Training data sources · Training · Lineage · LeRobot ACT: Kite ML, Inc.
Official LeRobot guidance uses the ACT policy with the SO-101 end-to-end workflow.
Evidence concerns the SO-100 configuration.
Full parameters and verification notes (model card)
Kite ML, Inc. made this task-trained derivative of Hugging Face LeRobot ACT. It is not a new foundation algorithm; Trossen Robotics provides the original recordings and robot, not this derivative policy. It is distinct from baseline 30,000 steps, all-seven 90,000 steps and real-recording 30,000 steps. No particular pretrained ACT checkpoint is established as its initialization.
Input: four RGB views (low, high, left wrist, right wrist), each 3×480×640, and 14 joint positions. Output: 14 joint targets in 100-step chunks; n_obs_steps=1, chunk_size=n_action_steps=100. It does not read language instructions. Configuration: ResNet18 with ImageNet1K V1 weights, Transformer width 512, 8 heads, feedforward 3200, ReLU, encoder 4 / decoder 1; VAE latent 32 and 4 VAE encoder layers, dropout 0.1, KL weight 10. MEAN_STD normalization for visual/state/action. Saved settings: CUDA, AMP/PEFT off, learning rate 1e-5, backbone learning rate 1e-5, weight decay 1e-4. These are not minimum inference requirements.
Training: kiteml/trossen-bin-picking-sim-baseline, 17 cup episodes, 6,867 frames; 90,000 steps, batch 8, one A100, about 4 hours, as reported by the author. LeRobot >=0.6 is the stated loader. The pinned experiment repository contains README, videos and results, without full training or rollout scripts; exact upstream training revision and full reproducibility remain unverified. The upstream LeRobot repository is implementation context, not proof that this exact training pipeline is published.
Author-only closed-loop MuJoCo evaluation: room light 8/20, low light 3/20, warm light 5/20, extra objects 10/20; total 26/80 (32.5%). Two held-out layouts from source episodes 4 (left pickup) and 14 (right pickup), 10 starts each per condition, perturbations within 1 cm. Placed means lifted at least 3 cm and finally inside the tray, max 19 s, complete 100-step plans executed. All successes were on the left layout; none on the right. Images are rendered by the simulator; the low view includes a real-station photo and cup textures. Wrist-camera cables now follow dataset shape and arm joints instead of fixed hooks. Evaluation protocols changed, so old/new results cannot directly measure model improvement. No real-robot run, independent reproduction or generalization proof is claimed.
Complete weights opened 2026-11-17 23:38:54 UTC (2026-11-18 07:38:54 Taipei), not the empty repository at 23:38:31 UTC. Public and nongated. model.safetensors: 206,765,304 bytes. We downloaded the complete file and verified SHA256 c2e798686cb00f6ff007285ce3f2907e4a92844f8fbf8fde333d7c55226941b8 against the LFS declaration, without running the policy. Configuration, pre/postprocessing and normalization files are present. Pin revision 353a961fb1c17a2f00b95fcbac4994282faf8eeb when calling ACTPolicy.from_pretrained.
Model-card metadata declares Apache-2.0; config.license is null and the repository has no standalone LICENSE. This is a publisher declaration, not a file-by-file legal audit; commercial permission stays unknown. Action units/control frequency, full-model trainable parameter count and minimum VRAM are unverified.