Figure AI Pushes robotics LLM physical AI 2026 Forward With New Training Methods

Figure AI released new training results last month that show robots handling objects they have never seen before. The work relies on large language models to supply basic reasoning during physical tasks.

The tests took place at the company's Bay Area lab. Robots picked up unfamiliar tools and adjusted grips without new programming. Engineers credit the addition of vision-language models for the change.

This development affects multiple teams. Boston Dynamics continues work on Atlas with its own control stack. Physical Intelligence focuses on data pipelines that feed models with real robot trials. Each approach now faces pressure to prove that LLM integration delivers reliable gains.

Figure AI and Physical Intelligence both report higher success rates on novel tasks when language models guide low-level actions.

Training data now includes language cues for object properties

Figure AI collected video of thousands of household and factory interactions. Each clip carries captions that describe weight, texture, and stability. The language model reads these descriptions and suggests grip adjustments before the robot moves.

Physical Intelligence runs a similar loop at smaller scale. Their system pairs short language instructions with force sensor readings. Early results show fewer dropped items on first attempts with new parts.

Boston Dynamics has shared less detail on language integration. Their recent updates still emphasize hardware balance and joint control.

Why these experiments matter beyond demos

Most robot failures happen when conditions differ from training data. Adding language models lets systems draw on broad knowledge of how objects behave. A model can infer that a glass cup breaks under pressure even if the robot never handled that exact cup.

The shift reduces the need for exhaustive physical trials. Companies still collect robot data, yet the model supplies starting heuristics that shorten the search.

Investors have noticed the change. Figure AI closed a new round this spring with participation from existing backers. Physical Intelligence announced partnerships with two auto parts suppliers for pilot lines.

Limits remain on speed and safety guarantees

Language models add steps between perception and action. Latency can reach several hundred milliseconds on current hardware. That delay matters on fast assembly lines.

Safety teams also note that model outputs stay probabilistic. A robot may receive conflicting suggestions when sensor noise or partial views occur. Companies now run secondary controllers that override language suggestions when force limits are reached.

Critics point out that public benchmarks still favor narrow tasks. Broad common-sense claims await longer trials in uncontrolled settings.

Next milestones to track over the coming months

Watch for Figure AI's planned public demo with a major logistics partner in August. The session will show the robot unpacking mixed boxes without pre-mapped layouts.

Physical Intelligence has said it will release a subset of its training logs by September. The data could let outside groups test their own model additions.

Boston Dynamics typically aligns major updates with its annual event in October. Any language model additions would likely appear there first.

These checkpoints will show whether the current gains hold when tasks grow more complex or when hardware moves to new sites.