Downloads · 30 days
0
jiankimr/pi05_libero10_ipec_seed42_step29999
pi05_libero10_ipec_seed42_step29999 is a robotics model from jiankimr. 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. The card lists the license as apache-2.0.
Clean (un-poisoned) pi0.5 (openpi) fine-tuned on LIBERO-10 only (IPEC-COMMUNITY/libero10nonoops1.0.0lerobot, 379 episodes / 101,469 frames @ 20 Hz). Used as the victim policy for fatigue-injection / defense-evasion ex…
Downloads · 30 days
0
Access
Public
Updated Sep 14, 2026
Repo size
12.4 GB
Likes
0
Public
Click a slice to open those files.
Other12.4 GB · 100%
From the Hugging Face model README
Clean (un-poisoned) pi0.5 (openpi) fine-tuned on
LIBERO-10 only (IPEC-COMMUNITY/libero_10_no_noops_1.0.0_lerobot, 379 episodes / 101,469 frames @ 20 Hz).
Used as the victim policy for fatigue-injection / defense-evasion experiments (LPF, jerk / COBALT jitter).
| Training seed | 42 |
| Step | 29999 (≈~76 epochs at batch 256) |
| LIBERO-10 success rate | 92% (10 tasks × 10 trials = 100 episodes, 500 Hz physics, OSC_POSE) |
| Base model | gs://openpi-assets/checkpoints/pi05_base |
| Recipe | openpi pi05_libero defaults: batch 256, FSDP over 2×H100, AdamW, LR warmup 10k -> 5e-5 then constant (no decay), EMA 0.999 |
| Action space | 7-D EE delta (pos xyz, axis-angle xyz, gripper), chunk 8 @ 20 Hz |
| Contents | params/ (orbax, EMA weights), assets/.../norm_stats.json, _CHECKPOINT_METADATA. train_state/ (optimizer) omitted (−31 GB). |
Why 10k: on this LIBERO-10-only dataset 10k steps ≈ 25 epochs, the same epoch budget as openpi's official 30k-step recipe on the 4-suite dataset (≈273k frames). SR plateaus from 10k (seed 42: 10k 94% / 20k 94% / 30k 92%).
The training config lives in the rsec_pi05 package (pi05_libero10_ipec: openpi's LeRobotLiberoDataConfig
re-keyed for the IPEC dataset). Download the repo to a directory CKPT and:
from openpi.policies import policy_config
from openpi.training import config as _config
import rsec_pi05.configs # registers pi05_libero10_ipec
cfg = _config.get_config("pi05_libero10_ipec")
policy = policy_config.create_trained_policy(cfg, "CKPT") # dir containing params/ and assets/
Sibling checkpoints: jiankimr/pi05_libero10_ipec_seed42_step10000, ..._seed42_step29999, ..._seed43_step10000
(seeds 44/45 to follow).
Training / evaluation / attack code, experiment settings and result summaries: https://github.com/jiankimr/h100vast-backup (private) - README there reproduces everything.