RoboTwin cross-attempt PIM checkpoint
The inference package is distributed separately from the code. Keep all files together:
cross-attempt-pim/
release.json relative paths, file hashes, memory mode
COMPLETE
README.md / NOTICE / LICENSE* loader guide and upstream weight terms
THIRD_PARTY_NOTICES.md
handoff/
checkpoint/pretrained_model/ final policy, processors, tokenizer
runtime/baseline/contract.json EEF16 / Joint14 / three-camera contract
reference/ frozen action mean and standard deviation
cte/ Mamba config and encoder weights
eap.safetensors EAP and PIM weights
memory/
task_memory.json
task_memory.safetensors train-only retrieval keys and values
retrieval.json
retrieval.safetensors
goal/ frozen language embedding and tokenizer
All model weights use Safetensors. The package contains no optimizer, training state, RNG dumps, training traces, videos, logs, or machine-specific paths. The separate frozen goal embedding preserves the original language-retrieval path without shipping another full foundation checkpoint. The policy contains Gemma-derived weights; keep the bundled Gemma terms and use restrictions with any redistribution.
Load
In a RoboTwin adapter with the matching LeRobot PI0.5 and Mamba runtime:
from openpi.zeva.pim_policy import ZevaCrossAttemptPIMPolicy
from huggingface_hub import snapshot_download
root = snapshot_download("chen123fu/Zeva-CrossAttempt-PIM")
policy = ZevaCrossAttemptPIMPolicy.from_release(root)