rskill-act-libero

Action Chunking Transformer (Zhao et al., 2023) fine-tuned on HuggingFaceVLA/libero, packaged as a OpenRAL rSkill for the LIBERO Franka-Panda embodiment.

Field	Value
Weights	hf://Deepkar/libero-test-act
Architecture	ResNet-18 backbone · 4+1 encoder/decoder · latent VAE · chunk_size=100
Inputs	image 256×256, image2 256×256, state (8-D)
Action	7-D delta-EEF + gripper
Robot	franka_panda (LIBERO embodiment tag)
Dataset	HuggingFaceVLA/libero (Apache-2.0)
License	Apache-2.0

Run

CC=/usr/bin/gcc uv sync --group libero      # first time only
ral sim run --config examples/sim/act_libero_spatial.yaml \
           --rskill rskill://rskills/act-libero

The shipped sim YAML pins libero_spatial/0 for a 200-step single-episode rollout. Sweep tasks with --task libero_spatial/<n>. A spot-check on libero_spatial/2 reaches is_success=True in ~91 steps (reward 1.0) on a single seed.

Camera & state contract

LIBERO emits images={"camera1": agentview, "camera2": eye_in_hand} while this checkpoint's input features are observation.images.image / observation.images.image2. The manifest's image_preprocessing block rewrites the batch keys at step time:

image_preprocessing:
  flip_180: true            # HuggingFaceVLA/libero is captured rotated 180°
  aliases:
    camera1: image
    camera2: image2

The state_contract.dim: 8 declaration confirms the proprio width. Because the upstream training set is HuggingFaceVLA/libero — the same dataset the smolvla / pi05 / xvla LIBERO checkpoints in this repo were finetuned on — the state semantics (pos3 + axisangle3 + grip2) line up with OpenRAL's LIBERO backend end-to-end, with no quat-vs-axisangle mismatch.

Benchmarks

None measured yet. Populate eval/ with ral benchmark run JSON fixtures before publishing a headline number.