Polaris.AI derivative of π0.5 for UR7e and RYODEN onsite preparation. Checkpoint 020000 includes processors, normalization and tokenizer assets. Public whole-repository download; derivative license unknown. Only configuration loading is reported; inference and robot rollout remain untested.
기능 분류 · 기록 유형 · Six joint positions and one gripper value; 70-step chunk, 30 action steps executed per configuration. · continuous · 계보 · π₀.₅ · 계보 · polarisai-robots/bento_ur7e_pi05_v5_color200@153d92de8799029b381508e563931395b1163747: Polaris.AI Robotics
Unitree 공식 G1 학습 안내는 Pi0.5 베이스 정책을 구성합니다. · 표 II의 PGX 전체 성공률(C0/C1)은 상자 운반 90%/60%, 컵 놓기 75%/90%, 커피 가져오기 60%/55%다. · 옥스퍼드 팀은 실제 시연 44건(조작 구간 176개)으로 30 Hz에서 π₀.₅ 정책을 미세조정했다. 그림 7의 잡기, 놓기, 다시 잡기, 전달 완료율은 각 단계에 도달한 회차를 기준으로 계산했다.
문서에 명시된 Unitree LeRobot 환경과 G1 데이터셋/구성이 필요합니다.
Isaac Lab Arena에서 Unitree G1과 Dex-3 손, HOMIE 전신 제어를 사용해 50 Hz로 학습 및 실행했다. 조건당 최소 20회 평가했다. C0는 표준 시작 자세이며 C1은 두 이동축에 각각 ±0.30 m, 요각에 ±30°의 무작위 편차를 적용한다. 상자 운반, 컵 놓기, 커피 가져오기의 제한 시간은 각각 40초, 60초, 110초다.
실제 정책과 50 Hz 시뮬레이션 정책은 별도로 학습했다. 그림 7은 시뮬레이션의 표준 시작 자세 C0와 비교한다.
SiMa.ai가 SO-101 팔로어 암으로 회색 지우개를 흰 바구니에 넣는 Pi0.5 정책의 배포 패키지를 업데이트했다. Modalix DevKit용 새 버전은 Palette SDK와 Model Compiler 3.0.0, LLiMa 0.6.0을 사용하며 컴파일 그래프가 38개에서 64개로 늘었다.
카메라 시점은 학습 조건과 일치해야 한다. PoLiMa와 Palette SDK 3.0.0 및 LLiMa 0.6.0을 실행하는 Modalix DevKit이 필요하다.
전체 매개변수 및 검증 설명 (모델 카드)
Publisher: Polaris.AI Robotics, an HF company account linked to polarisai.co.jp (Polaris.AI Inc.). Upstream algorithm: Physical Intelligence π0.5. This is a derivative artifact, not a new official upstream release.
Artifact: polarisai-robots/bento_ur7e_pi05_nextech_v1; checkpoint 020000. First weight upload: 2026-11-17T07:54:28Z, commit 6eb160def0c52d56915f250a0d76fb283b127530. Documentation revision: 68172d792d82e01a53c28ce564908522c7427f35 at 07:55:05Z. Empty repository creation at 07:41 is not weight availability; original training completion time unknown.
Initial model: polarisai-robots/bento_ur7e_pi05_v5_color200@153d92de8799029b381508e563931395b1163747. Publisher reports no retraining or conversion during upload. Training updates 20000; batch16; learning rate0.000025; train_expert_only=false; freeze_vision_encoder=false.
Safetensors weight file:9354050784 bytes. HF-reported SHA256:f30b147ceba46fd892621ec44a1ad28f48546b80f0e08e2bc87b14052f4e183d, not independently recomputed. HF metadata counts4143404816 parameters:533550864 F32 and3609853952 BF16. Config dtype=bfloat16 does not imply all tensors are BF16.
Configured visual features:base_0_rgb,left_wrist_0_rgb,right_wrist_0_rgb, each3×224×224 RGB. Internal state width32; output action7; chunk70; n_action_steps30; n_obs_steps1; num_inference_steps10; max_action_dim32 is internal padding width. Actions:joint_0.pos through joint_5.pos plus gripper.pos.
The card describes six joints plus gripper and top/wrist video in local data, not three verified physical cameras or32 joints. Mapping:top→base_0_rgb;wrist→left_wrist_0_rgb. empty_cameras=0; missing-camera processing remains unverified. STATE/ACTION use QUANTILES; VISUAL uses IDENTITY. Pre/postprocessors, normalization files and tokenizer must accompany weights. use_peft=false;use_amp=false;use_relative_actions=false.
Recorded training dataset identifier:polarisai-robots/bento_ur7e_nextech_v1. Exact training dataset identity/count unknown:discussion40 episodes conflicts with local42 episodes,80785 frames,60fps. Do not treat local counts as training counts or60fps as control frequency. No private discussions or intranet were accessed.
Runtime reported by card:audited plugin main b492a4e7f02c51417c1ed58f2250f6df94139d0c, but public GitHub API returned404 during this review; source availability unconfirmed. Configuration-loading environment:LeRobot0.6.2/core a4b2744b24415b01e6deb4f2638971de125dde33/plugin5ba2bbdc860a4741f4d44765b8748a3d1a1d2fdf. Setup defaults to core v0.6.0, so plugin branch alone does not specify a tested runtime.
Only PI05Config.from_pretrained was reported successful. GPU weight loading,inference,robot rollout and customer-PC compatibility were not tested. A6000/RTX PRO6000 are candidates,not minimum hardware requirements. Select the new repository explicitly:original config.repo_id still names the initial model. For continued training choose a new output directory and control Hub publication.
Public non-gated weight download; weight_license=unknown; commercial_use=unknown. No license in card/file tree;config.license=null. Public access and upstream licensing do not authorize derivative commercial use; upstream terms require separate review. API availability unknown.