SLIM-LIBERO

SLIM for LIBERO

This repository contains the released SLIM Stage 2 policy checkpoint for LIBERO and LIBERO-Plus. SLIM is a compact latent interaction policy for robot manipulation.

Checkpoint

• Stage 1: action-grounded masked trajectory prediction on LIBERO all+90
• Stage 1 objective: IDM:FDM = 0.125:1 for 3 epochs
• Stage 2: flow-matching policy training on LIBERO all for 40 epochs
• Stage 1 and Stage 2 video backend: torchvision_av
• Stage 1 EMA: enabled, momentum 0.999
• Stage 2 EMA: disabled
• Action horizon and execution chunk: 8
• Image size: 224 x 224, agent and wrist views
• State/action dimensions: 7/7

The checkpoint is a plain PyTorch state_dict and loads directly with SLIM.

Results

LIBERO suite scores are 94.40% (LIBERO-10), 99.40% (Spatial), 99.40% (Object), and 96.80% (Goal). The complete LIBERO-Plus suite/category reports are included under evaluation/.

Usage

Install SLIM and configure the DINOv2 and T5 paths as described in the SLIM README. Then run a policy server from the SLIM repository root:

python -m slim.serving.server \
  --checkpoint /path/to/SLIM-LIBERO/checkpoints/epoch_40_pytorch_model.pt \
  --port 10093 \
  --bf16

The checkpoint requires the included config.yaml and action_stats.json to remain in the repository root. See checkpoint_manifest.json for hashes and the exact release revision.

Limitations

This checkpoint is intended for research evaluation in LIBERO-compatible simulation environments. It should not be deployed on physical robots without task-specific safety validation and action-bound checks.