Kite task-trained ACT weights for cup picking on a bimanual WidowX AI station, based on LeRobot ACT; 30,000 training steps, evaluated by the author in simulation only.
Single cup task; no language input. Author simulation only, all successes on the left layout, no hardware deployment reported. Exact upstream training revision and action units/frame/frequency remain unverified. Verfügbare Gewichte oder Code belegen keine sofortige Roboterkompatibilität. Prüfen Sie Konfiguration, Daten und Schnittstellen in der Quelldokumentation.
Modalix DevKit, PoLiMa mit SDK-3.0-Unterstützung und ein SO-101-Follower mit USB-Kameras über dem Arbeitsbereich und am Handgelenk, passend zu den Trainingsansichten.
SO-100-Followerarm in Forschung und Evaluation; keine Aussage zur Werksvorinstallation.
Alle Parameter und Prüfhinweise (Modellkarte)
Kite ML, Inc. publishes this task-trained derivative of Hugging Face LeRobot ACT, not a new foundation algorithm. No fine-tuning from a specific pretrained ACT checkpoint is established. The upstream implementation revision used for training is undisclosed; the model card states LeRobot >=0.6 can load it.
The task is to move the red cup into the clear tray on a bimanual WidowX AI station. Inputs: four RGB cameras (low, high, left wrist, right wrist), each 3×480×640, plus 14 joint positions. Outputs: 14 joint targets per step, chunks of 100 steps; n_obs_steps=1 and n_action_steps=100. It does not consume language instructions.
Configuration: ResNet18 with ImageNet1K V1 backbone weights; Transformer width 512, 8 heads, feedforward width 3200, ReLU, 4 encoder and 1 decoder layers; VAE enabled, latent dimension 32, 4 VAE encoder layers; dropout 0.1, KL weight 10. Visual, state and action normalization use MEAN_STD. Training config has CUDA, no AMP or PEFT, optimizer learning rate 1e-5, weight decay 1e-4 and backbone learning rate 1e-5. These are saved settings, not measured deployment requirements.
Training uses 18 cup episodes from Trossen public real bin_picking recordings, with source episodes 4 and 14 held out, totaling 7,318 training frames. Batch 8 on one A100, 30,000 steps, about 1h50m. Training hardware is not a minimum inference requirement.
Author-reported closed-loop simulated-station placements: room light 10/20, low light 8/20, warm light 8/20, extra objects 7/20, total 33/80 (41.25%). Every placement was on the left-arm pickup layout; none succeeded on the right-arm layout. The low camera combines a real-station photo with cup cut-outs. This is simulation only, not a hardware success rate or an independent reproduction. The website's 91% example is unrelated.
Complete weights first uploaded on 2026-11-16 at 22:39:44 UTC (2026-11-17 06:39:44 Taipei). The initial empty repository at 22:39:02 is not the complete-weight opening event. model.safetensors is 206,765,304 bytes; config and pre/postprocessor normalization files are included. Public, nongated HF download; pin the exact commit below. Model-card metadata declares Apache-2.0, but there is no standalone LICENSE file and config.license is null. This records the publisher's declaration, not a file-by-file legal audit; commercial permission is left unknown. LeRobot's code license is a separate claim. Full-model parameter count, action coordinate frame/units, deployment control frequency and independent/hardware evaluation are not established.