Config for the BlindVLA LeRobot dataset.

Form global batches, then split evenly across ranks.

ZMQ ROUTER server that handles multiple concurrent clients. Uses a ROUTER socket to accept connections from multiple clients simultaneously.

Minimal policy interface required by training/eval/serving scripts.

Minimal LeRobot dataset wrapper for Bbox Cotrain - Keeps actions as-is (no extra transforms). - Emits processed keys defined in `vla_scr

Minimal LeRobot dataset wrapper for BlindVLA. - Keeps actions as-is (no extra transforms). - Emits processed keys defined in `vla_scratc

Single-sample dataset providing the fields expected by ToDataSample.

Create a fresh iterator each epoch without prefetching.

Invert network action targets back to dataset action format. Expects `sample["actions"]` to contain per-step targets of shape [K, 10]: [Δpo

Convert Libero actions into relative 6D pose deltas.

Convert history of Libero states into global 6D pose.

Stack Libero camera streams into a standard tensor layout.

Convert history of Libero states into relative 6D pose deltas.

Enable on-screen rendering while keeping offscreen buffers for images.

Prepare image + prompt inputs for PaliGemma VLM bridges.

Prefetch the next epoch's first batch while the current epoch runs.

Tokenize prompt using Qwen3-VL chat template and produce policy inputs.

Tokenize prompt using SmolVLM chat template and produce policy inputs.