Toyota announces large behavior models
Toyota Research Institute reported that it had taught robots more than 60 difficult, dexterous skills with a method based on Diffusion Policy, and named the direction large behavior models, by analogy with large language models.
Why it matters
Here an industrial laboratory, rather than a university group, put its name to the claim that a robot’s skills would come from one trained model rather than from per-task engineering, and gave a skill count for it. The term itself stuck from this point.
The skills include pouring liquids, using tools and manipulating deformable objects. A behaviour is learned from a teacher’s haptic demonstrations together with a language description of the goal, after which a diffusion policy absorbs what was shown; a new behaviour can be put into service from dozens of demonstrations. Because the skills work directly from camera images and tactile sensing, using only learned representations, they cope with tasks that were traditionally very hard: deformable objects, cloth and liquids. The figure of 60 is TRI’s own count: the press release gives no per-skill success rates. TechCrunch’s report of the company’s presentation of 19 September 2023 quotes Russ Tedrake, TRI’s vice president of robotics research, saying 60 skills were trained already and 100 were expected by the end of the year, and gives some success rates (pancake flipping 90 per cent, 27 of 30; dough rolling 96 per cent; food serving 90 per cent). Those too are TRI’s own, with no independent check, so the confidence on this record stays medium.