Kite’s 30,000-step baseline ACT derivative for red-cup picking on a bimanual WidowX AI station. Four RGB views and 14 joint states produce 100-step action chunks. Author simulation evaluation only.
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. authored this task-trained derivative of Hugging Face LeRobot ACT: kiteml/trossen-bin-picking-sim-act-cup-baseline, baseline, 30,000 training steps. Trossen Robotics provides the original robot/data, not this derivative's developer. Separate from 90k, all-seven and real-recording policies; no specific pretrained ACT checkpoint or exact upstream training revision is established.
Inputs: four RGB cameras (low, high, left wrist, right wrist), 3×480×640 each, plus 14 joint positions. Output: 14 joint targets per step, full 100-step chunks; n_obs_steps=1, n_action_steps=100. No language input. LeRobot >=0.6. ResNet18 (ImageNet1K V1), Transformer width512, 8 heads, feedforward3200, encoder4/decoder1; VAE latent32, VAE encoder4, dropout0.1, KL10, MEAN_STD normalization. Saved config is not a minimum inference specification.
Training: 17 cup training episodes from kiteml/trossen-bin-picking-sim-baseline, 6,867 frames; batch8, one A100, 30,000 steps, about78 minutes. Simulated data derived from Trossen public bin_picking recordings. Training hardware is not a minimum deployment requirement.
Evaluation update, not new weights or algorithm: model-card commit 6920362f97f9134fc708510fb8261895b102ba28, 2026-11-17 23:39:04 UTC; report commit f61bcdc1211db23873c28aa785cb4ada8ccccaad, 23:39:14 UTC (Taipei 2026-11-18 07:39:14). Protocol is the report revision identified by this commit, not an invented publisher version number. MuJoCo closed-loop simulated station, two held-out layouts: source episode4 left arm and14 right arm; 20 trials per condition,10 per layout, starting cup perturbation within1cm. Success means lift at least3cm and finish inside tray; up to19s, executing full100-step action chunks. Wrist-camera cables now follow dataset shapes/joints rather than fixed hooks. Test images use the simulator renderer; low camera includes a real station photograph and cup cut-outs. This creates training/test image-domain mismatch.
Current author-reported results: room light5/20, low light0/20, warm light0/20, extra objects7/20; total12/80 for this protocol only. All successes on left-arm layout; no right-arm success. Historical fixed-hook protocol at card f962e784be412dde1c3848edd0d7081bd5c24bed: room light8/20, low light2/20, warm light0/20, extra objects8/20, total18/80. Preserve history separately; changed protocol prevents interpreting the change as model capability improvement or decline, pooling trials or writing a generic success rate. No hardware test, independent reproduction or proof of generalization.
Original full weights uploaded2026-11-08 23:04:21 UTC (Taipei11-09 07:04:21), commit c3fed118029dc8a8baeab43aee3d9a25ccc451f5; this is historical release identity, not this event's launch. Old/new file trees differ only in README.md; model.safetensors LFS object and byte size are identical. Model-card metadata declares Apache-2.0; no file-by-file legal audit, independent LICENSE absent, config.license=null. Commercial-use permission remains unknown.
Sources checked2026-11-18. Exact action units/coordinate frame/control frequency and company financing remain unverified.
Hugging Face metadata for the pinned original weight revision reports 51,685,006 parameters (F32); recorded as hosting metadata, not an independent tensor enumeration by this assistant.