focus-vlwa

focus-vlwa checkpoint

This checkpoint supports head_history inference for RoboDojo arx_x5 joint-action evaluation. It uses 20 full head-camera history frames, a 14-value joint observation, 32 internal action dimensions, and a 50-action prediction horizon.

Inference files

Keep the following files together:

focus-vlwa-release/
  README.md
  model.safetensors
  model_config.json
  assets/arx_x5_sim/norm_stats.json
  SHA256SUMS

This inference package excludes training metadata and optimizer state. The PaliGemma tokenizer is resolved separately; provide tokenizer_path for offline evaluation.

Usage

Install the matching Focus-VLWA code and its Transformers patch. Use this directory as the checkpoint path:

export CHECKPOINT=/lpai/volumes/base-eb-ali-wl/jingxie/openpi-assets/checkpoints/focus-vlwa-release

For local inference, construct FocusVLWAPolicy(checkpoint_path, asset_id="arx_x5_sim", num_steps=10). For XPolicyLab/policy/FocusVLWA, pass "$CHECKPOINT" as the checkpoint argument and use env_cfg_type=arx_x5, action_type=joint, and history_mode=head_history. The adapter records head-camera observations at every control step, samples history at 25-step intervals, and resets history between episodes. Evaluation uses 10 denoising steps and executes 25 of the 50 predicted actions before replanning.

File integrity

Run from the release directory:

sha256sum -c SHA256SUMS

RoboDojo simulation assets, the tokenizer, the matching Focus-VLWA code, and XPolicyLab are separate dependencies.