이미지: METAL LAB 생성
Summary
- Sakana AI has expanded its RSI Lab and named physical AI as its next research frontier
- The goal is to build world models that let agentic reasoning systems improve their own ability to simulate, plan, and act in the real world
- No specific model names, investment size, or launch timeline have been disclosed
- 발표 주체
- Sakana AI
- 발표 일시
- 2026-08-10(UTC)
- 채널
- X(트위터) 공식 계정
- 핵심 주장
- 물리 AI(Physical AI)가 재귀적 자기개선(RSI)의 다음 프론티어
- 조직 변화
- RSI Lab 확장
- 연구 목표
- 월드모델을 통해 에이전트 추론 시스템의 시뮬레이션·계획·실세계 행동 능력을 재귀적으로 개선
- 부연 언급
- "일본이 로보틱스 혁명을 촉발했다"는 취지의 발언
An AI company named after a fish turns toward robots
On August 10, 2026, Sakana AI announced via its social media that it will make robotics and real-world AI — "physical AI" — the company's next research axis. The company said it is expanding its existing RSI Lab to build world models that let agentic reasoning systems improve their own ability to simulate, plan, and act in the real world. In accompanying remarks, the company stated that "Japan sparked the robotics revolution."
Since the news was disclosed via a single tweet, details such as specific model names, investment scale, or launch timing have not yet been revealed. Still, the direction in which the company intends to reorganize its research organization comes through fairly clearly.
What is Sakana AI
Sakana AI is a startup founded in Tokyo in 2023, reportedly by researchers who came from Google Brain. The company's name, "Sakana" (魚), means fish — a nod to its philosophy of having many small models cooperate like a school of fish to solve problems, rather than scaling up a single massive model. This approach aligns with the company's long-standing emphasis on evolutionary algorithms and nature-inspired learning methods.
RSI: AI that fixes itself
Recursive Self-Improvement (RSI) refers to a process in which an AI system evaluates its own capabilities and revises its own learning methods or architecture. The idea is that the system repeats improvement loops by identifying its own weaknesses, rather than relying on humans to manually design new data or new architectures each time. Until now, this discussion has mostly taken place in the domain of language models handling text and code. The targets of evaluation and improvement have largely stayed within the digital realm — raising benchmark scores or improving code-review accuracy.
The physical AI that Sakana AI is describing proposes extending this loop into the real world. The idea is to absorb data generated as robots grasp objects, walk, and react to unexpected friction or vibration back into training, jointly improving both simulation and real-world action capabilities.
What a world model does
A world model internally predicts how an environment will change. Such models are used, for example, to let a self-driving car predict road conditions in the next moment, or to calculate what posture a robotic arm should take to regrasp an object it has dropped. Sakana AI appears to be aiming to combine this world model with an agentic reasoning system, creating a cycle in which robots rehearse countless scenarios in virtual simulation beforehand and then validate them in the real environment.
| Category | Self-improvement in existing language models | Self-improvement in physical AI |
|---|---|---|
| Training data | Mainly text and code | Simulation and real sensor data |
| Evaluation criteria | Benchmark scores | Success rate and safety of robot actions |
| Representative challenge | Hallucination (generating factually incorrect answers) | Violation of physical laws, sensor noise |
Why robotics
Japan is generally regarded as a country with a long-accumulated technological base in industrial robots and automation equipment. Sakana AI's framing of itself as an extension of Japan's robotics tradition appears related to this background. The term "physical AI" has also been increasingly used across the global AI industry recently, reflecting a broader trend in which the term is becoming shorthand for AI that operates in physical space, beyond language and images.
So what changes
This announcement is closer to a declaration of research direction than a finished product or completed model. Still, it can be read as a signal that the discussion of "self-improvement," which had been concentrated around language models, is expanding into the domain of robots' physical actions. As the center of gravity in the competition shifts from AI that writes text well to AI that handles objects well, Sakana AI has added its name to that trend. What concrete results emerge, and in what form and timeline, will have to be confirmed through future disclosures.
