Vision-Language-Action Models

• From vision-language to action: grounding language instructions in physical actions
• VLAs: architecture (vision encoder + LLM + action head), RT-2, Octo, OpenVLA
• Action tokenisation: discretising continuous actions, action chunking
• Pretraining recipes: web-scale vision-language data → robot manipulation data
• Generalisation: unseen objects, environments, instructions
• Co-training with internet data and robot data
• Embodiment-agnostic models: one model for multiple robot form factors
• Benchmarks: SIMPLER, real-world evaluation protocols