TurboVLA：LLM在VLA中必须是核心吗？消费级显卡实现32Hz实时操作

TL;DR - TurboVLA (Huazhong University of Science and Technology + Huawei) is a vision-language-action robot policy that removes the LLM backbone from the per-step execution path, running a full 0.2B policy at ~32Hz on a single RTX 4090 with 0.9GB VRAM. It matters because real-time, low-cost local inference — not just task success rate — is what makes VLA policies deployable next to the robot.

• Replaces the conventional V→L→A path with V+L→A: DINOv3 visual features and a lightweight text encoder (e.g. BERT) meet in a bidirectional vision-language interaction module, then fuse with robot state into a lightweight decoder that predicts continuous action chunks in parallel.
• LIBERO (40 language-conditioned tasks, 2,000 evals): 97.7% average success at 0.2B params, 0.9GB inference VRAM, 31.2ms end-to-end latency vs π0.5 at 96.9% and 93.6ms (~11Hz) on the same GPU — ~6% of the parameters.
• Language is trimmed, not removed: ablation shows removing language drops LIBERO average success from 97.7% to 70.8%, and natural-language instructions beat Task-ID conditioning.
• Scales beyond single-arm: 0.4B config hits 60.2% on RoboTwin 2.0 dual-arm (14-dim) clean setting at 43.4ms, above π0.5 (57.0%) and StarVLA-α (50.3%); real AgileX Piper trials on four tasks reached 92.5%/80%/90%/87.5%, each above π0.5.