
로봇 팔
SO-101
관련 버전 2
Modalix DevKit, SDK 3.0을 지원하는 PoLiMa, 학습 시점과 일치하는 상단 및 손목 USB 카메라를 장착한 SO-101 팔로워.
Kite's task-trained LeRobot ACT policy for cup picking, trained on seven simulated datasets for 30,000 steps. This records the author's revised simulation evaluation, with unchanged weights.
전체 모델 설명 보기논문 / 보고서 미확인
공식 소개 자료가 등록되어 있으며 해당 버전의 논문이나 기술 보고서는 아직 확인되지 않았습니다.
최근 확인: 2026-11-18
자료 게시일: 2026-11-17
| 파일 | 접근 방법 | 가중치 라이선스 |
|---|---|---|
| trossen-bin-picking-sim-act-cup-all · 30000 | 접근 방법 | 가중치 라이선스 |
출처확인 날짜 · 2026-11-18
Author simulation only; two layouts; changed protocol and render-domain bias; no hardware or independent reproduction. 가중치나 코드가 공개되어도 로봇에서 즉시 실행할 수 있다는 의미는 아닙니다. 문서에서 구성, 데이터 및 인터페이스 요구 사항을 확인하세요.
joint_position · continuous
14 DoF · horizon 100 · mean_std
Hugging Face LeRobot ACT
단일 본체
Action control
공식 가중치
행동
robotics
chunk_size=100; n_action_steps=100
양팔 플랫폼 / 로봇 팔
단일 작업
48,069 training frames
Seven simulated datasets; 119 cup episodes
모방 학습
파생 모델: ACT · LeRobot ACT ↗
Kite ML, Inc. publishes kiteml/trossen-bin-picking-sim-act-cup-all: the all-seven, 30,000-step task-trained derivative of Hugging Face LeRobot ACT. It is not a new ACT algorithm or a new weight release on 2026-11-18. No initialization from a specific pretrained ACT checkpoint is established; the exact upstream training revision is undisclosed. Trossen Robotics supplies the original recordings/robot platform, not authorship of this derivative. Other baseline, 90k and cup-real artifacts remain separate.
Training: 17 cup episodes in each of seven simulated datasets (baseline, dim, very-dim, warm, fluorescent, window, extra-objects), 119 episodes / 48,069 frames. Batch 8, one A100, 30,000 steps, about 78 minutes. Training hardware is not a minimum inference requirement. Four RGB inputs (low, high, left and right wrists), each 3×480×640, plus 14 joint positions; outputs 14 joint targets in 100-step chunks, n_obs_steps=1, n_action_steps=100. No language input. Bimanual WidowX AI context, LeRobot >=0.6.
Saved configuration: ResNet18 / ImageNet1K V1; Transformer width 512, 8 heads, feedforward 3200, ReLU, encoder 4 / decoder 1; VAE enabled, latent 32, VAE encoder 4; dropout 0.1, KL 10; MEAN_STD for images/state/action; CUDA device; AMP and PEFT disabled; learning rate 0.00001, weight decay 0.0001, backbone learning rate 0.00001. These settings are unchanged in this event.
Evaluation protocol: author-run closed-loop MuJoCo simulated station, protocol identified by report commit f61bcdc1211db23873c28aa785cb4ada8ccccaad (2026-11-17 23:39:14 UTC; Taipei 2026-11-18 07:39:14). The model-card update is 2a30bac7845f0f30b32fd87a73044aced5e3ab57, 2026-11-17 23:39:05 UTC. Sources checked 2026-11-18. Held-out source episode 4 is left-arm pickup, episode 14 right-arm pickup; each condition has 20 trials (10 per layout), cup starts within 1 cm. Success requires lifting at least 3 cm and ending inside the tray. Maximum 19 seconds, executing each 100-step chunk in full.
Current author-reported placements: room light 2/20; low light 0/20; warm light 3/20; extra objects 1/20; total 6/80 for these four conditions only. All six placements are on the left-arm layout; right-arm layout 0/40. Low light is 30–60%, warm light 2700–3500 K. Wrist-camera cables now follow the datasets' shape and arm joints, replacing fixed hooks. Test images are simulator-rendered; the low camera uses an empty real-station photograph and real-cup cutouts. This differs from the training render domain and is not evidence of generalization.
Historical protocol only: old card e2ccca8fc27cad57d2ac308970a1d023b470c847 reported room 2/20, low 2/20, warm 1/20, extra objects 0/20 with the older fixed-hook simulation. Preserve separately; never pool with current trials or interpret the change as a weight/capability gain or loss. No real-robot test or independent reproduction. These are publisher claims checked against its report and run JSON, not independently validated outcomes.
Complete weights originally uploaded 2026-11-08 23:04:42 UTC (Taipei 2026-11-09 07:04:42), commit 991c82d5d7e8eff8fd0851c1b98446b3211a4545. Only README.md changed between old and new trees. model.safetensors remains 206,765,304 bytes, LFS SHA256 51ee9bb19ccea385c3a66e83cd2c63210c38cf410081bedfe387392be73d1833. The current pinned tree is a metadata revision of that existing checkpoint. Public weights and datasets; model-card metadata declares Apache-2.0, config.license=null, no standalone LICENSE in the tree. This is the publisher's declaration, not a file-by-file legal audit. Commercial permission, full-model parameter count, action frame/units and minimum inference hardware remain unverified.