Kite task-trained LeRobot ACT artifact for red-cup picking on bimanual WidowX AI, trained for 90,000 steps on all seven simulated datasets. This event updates the simulated evaluation protocol and author-reported results; weights are unchanged.
Author simulation only; no hardware test or independent reproduction. Protocols and image domains differ. All successes on the left layout. 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 algorithm. Fine-tuning from a specific pretrained policy is unestablished. Trossen Robotics supplies the original recordings and robot, not development of this derivative. Exact artifact: kiteml/trossen-bin-picking-sim-act-cup-all-90k (all-seven 90k), separate from baseline, 30k and cup-real.
Training: 17 cup episodes from each of seven simulated datasets, 119 episodes and 48,069 frames; batch 8 on one A100, 90,000 steps, about four hours. Inputs: four RGB views (low/high/left wrist/right wrist), each 3×480×640, plus 14 joint positions. Outputs: 14 joint targets, 100-step chunks; n_obs_steps=1, n_action_steps=100; no language input. LeRobot >=0.6. Saved config: ResNet18/ImageNet1K V1; Transformer width 512, 8 heads, feedforward 3200, encoder 4/decoder 1 layers; VAE latent 32 and 4 encoder layers, dropout 0.1, KL weight 10, MEAN_STD normalization; CUDA, AMP/PEFT off, learning rate 1e-5, weight decay 1e-4, backbone learning rate 1e-5. These are training settings, not minimum inference requirements.
Evaluation protocol is identified by report commit f61bcdc1211db23873c28aa785cb4ada8ccccaad, dated 2026-11-17 23:39:14 UTC (Taipei Nov18 07:39:14). Model card 06191c312dcf4efbe7b5b2c719ae1c066f3718b4 is dated 23:39:06 UTC. Kite reran closed-loop evaluation in a MuJoCo station twin: two held-out layouts, source episode4/left arm and episode14/right arm; 20 trials per condition, 10 per layout, starts within 1 cm and shared across policies. Placement requires lifting at least 3 cm and ending in the tray; maximum 19 seconds, complete 100-step chunks. Wrist-camera cables now follow dataset curves and arm joints instead of fixed hooks. Test images use the simulator renderer; the low camera includes a real-station photo and cup cut-outs, differing from training images.
Current author-reported placements: room light 2/20, low light 3/20, warm light 1/20, extra objects 3/20, total 9/80 within this protocol only. All successes are on episode4/left arm; none on episode14/right arm. Low light 30–60%, warm light 2700–3500K. Author simulation report only; no independent reproduction, real-robot test or generalization proof.
Historical environment only: old card 73a9aedbda37c2cc236bc2aec29f79bd2806ce3a, fixed-hook cables: 8/20, 3/20, 4/20, 3/20 (18/80). Protocols changed; do not pool results or interpret the difference as a model capability change. No generic success-rate field is populated.
Complete weights first uploaded 2026-11-13 02:08:45 UTC (Taipei 10:08:45); this event is not a weight release. Only README.md differs between HF trees. model.safetensors remains 206,765,304 bytes, LFS SHA-256 dddec821ee28f1dfcad38361a36d4658c6968acca90de2905bb997387516a215. Checkpoint pins original weight commit 73a9aed; new evaluation resources pin the new card/report. Public nongated HF access. Apache-2.0 model-card declaration is unchanged, config.license=null; no file-by-file legal audit, commercial permission unknown. Full parameter count, exact upstream training revision, action units/frame and hardware control frequency remain unverified.