G0.5: One Autoregressive Stream for Robot Reasoning and Action

TL;DR - G0.5 is an autoregressive vision-language-action model that generates reasoning and robot actions from one transformer decoder. Its unified design enables prompt-steerable behavior and strong results across seven robotic evaluation regimes.

• Uses a learned tokenizer to represent actions from heterogeneous robots in a shared vocabulary.
• Interleaves task decomposition, object grounding, action hints, and action tokens in one chain-of-thought stream.
• Adds visual memory through the vision encoder to incorporate multi-second histories.
• Outperforms cited baselines across real-world fine-tuning, long-horizon manipulation, zero-shot transfer, and simulation benchmarks.