Downloads · 30 days
0
Shiki42/pi05-scan-object-uniform50-arm-active-mask-lora-10k-e251
pi05-scan-object-uniform50-arm-active-mask-lora-10k-e251 is a robotics model from Shiki42. Use it for the robotics task on the model card, and read the license before you ship it in a product. It is set up for openpi.
Do not reuse this checkpoint or treat evaluations derived from it as valid scientific results.
Downloads · 30 days
0
Access
Public
Updated Sep 19, 2026
Repo size
6.3 GB
Likes
0
Public
Click a slice to open those files.
Other6.3 GB · 100%
From the Hugging Face model README
Do not reuse this checkpoint or treat evaluations derived from it as valid scientific results.
废弃原因:当前 CTR IdleMask 的定义/实现存在问题,该固定 checkpoint 使用了问题 mask 训练。
CTR datasets are the intended scope for IdleMask. This deprecation concerns the incorrect mask definition/implementation used for this checkpoint; it does not recommend disabling IdleMask for CTR. Historical training completion, checksum verification, success rates and archive approvals are preserved for diagnosis only and do not restore eligibility.
Correcting labels or changing inference settings cannot repair these trained weights. An explicitly audited corrected dataset revision, newly trained checkpoint and separate evaluation are required. No weights have been removed or modified by this metadata update.
Owner-authorized deprecation dated 2026-09-19. This notice supersedes earlier reuse or validity statements below.
PI0.5 single-expert LoRA fine-tuning for scan_object, with 10,000 optimizer updates, batch size 16 and seed 87431. Initialized from the pinned PI0.5 Base JAX weights.
Dataset: Shiki42/ctr-scan-object-uniform50-20260917 at
c46453109ca9e99b4561faf34a73770644917a2b (50 episodes).
The published observation.arm_active_mask is used directly as left/right
supervision weight: 1 active, 0 masked, all seven action channels per arm.
Action horizons retain the upstream episode-padding mask. This masks waits
marked inactive by the published arm-active field; it does not use retime.*_idle.
The repository contains inference parameters, the matching normalization assets, resolved configuration, provenance and SHA-256 hashes. Optimizer and loader state are not published; this is not a resumable-training snapshot.
The paired evaluation is E252, using the historical frozen 100-scene scan_object suite. Evaluation results are recorded separately and must not be inferred from a successful upload. Results remain reported / audit pending until explicit archive approval.