Kini o ṣẹlẹ
Simate, a startup founded three months ago, released its first physical AI model called Simate‑beta. According to 36kr, the model demonstrated capabilities in memory, long‑horizon tasks, fine manipulation, and task adaptation, and achieved the top spot on the RoboDojo leaderboard with an average score of 33.95 (SR 27.96%). The company says the base model was not specially optimized for the leaderboard and that the result reflects its AI‑native research platform, AutoResearch, which automates hypothesis testing, training, and evaluation. The startup also announced that its research tools have been used by external researchers from MIT, Caltech, Tsinghua University and Peking University in a beta test, and that it has completed multiple financing rounds worth hundreds of millions of RMB. The team plans to release technical papers, open‑source components in phases, and to roll out further results by the end of 2026.
Simate announced Simate‑beta, describing it as a "general‑purpose physical fast system" that integrates a high‑level planning component (the "slow system") with a rapid perception‑action component (the "fast system"). The company claims the model can remember recent events, understand complex instructions, and execute millisecond‑level adjustments without relying on task‑by‑task retraining.
The company attributes its rapid progress to an AI‑native research called AutoResearch, which automates hypothesis generation, experiment execution, and result analysis across simulation and real‑robot environments. According to the report, this pipeline allowed Simate to top the RoboDojo leaderboard within three months of its founding.
External researchers from top universities have participated in a beta test of the AutoResearch platform, and the startup has secured several financing rounds totaling hundreds of millions of RMB. The team plans to publish three research papers and release parts of the platform as open source later in the year.
Simate‑beta’s leaderboard performance was reported on September 23, 2026, with an average score of 33.95 and a success rate of 27.96%. The company notes that the leaderboard evaluation does not fully capture real‑world robot performance, which will be disclosed separately.
Awọn alaye orisun: eu.36kr.com ↗
Kini idi ti o ṣe pataki
The launch marks a notable step toward general‑purpose physical AI that can operate in real‑world environments without task‑specific retraining. If the claimed performance holds, Simate‑beta could accelerate robot deployment in industries such as logistics, manufacturing, and home assistance, where high‑speed perception and long‑term planning are critical. The company’s AI‑native R&D workflow—combining a “SiPAI” model framework, an AutoResearch engine, and custom infrastructure—promises to dramatically increase research throughput, potentially reshaping how robotics labs iterate on models. However, independent verification of the leaderboard scores and real‑world performance is lacking, and details on model size, hardware requirements, pricing, and commercial availability remain undisclosed. These unknowns limit immediate practical impact but highlight a potentially transformative approach to robot AI development.
Physical AI that can generalize across tasks without extensive retraining is a long‑standing challenge in robotics. Simate‑beta’s claimed ability to combine long‑term memory with fast, precise motor control could reduce the engineering effort required to deploy robots in new settings.
The AutoResearch workflow, if effective, could serve as a template for other AI labs seeking to accelerate robot research, potentially lowering the barrier to entry for smaller teams and academic groups.
The involvement of researchers from institutions such as MIT and Tsinghua suggests early academic interest, which could foster collaborations and independent validation of the technology.
The lack of disclosed model size, compute requirements, and pricing means that the system’s accessibility to developers and enterprises remains uncertain. Without independent verification, the leaderboard claim alone does not guarantee real‑world efficacy.
Ibaraẹnisọrọ Mechanism: Bii O Ṣe Nṣiṣẹ Lootọ
Ṣawari imọ-ẹrọ abẹlẹ lẹhin idagbasoke yii ni ibaraenisọrọ.
crm_get_transaction(id='4092').In AI, what are a model's "parameters"?
Kini lati wo tókàn
Key areas to monitor include: (1) independent benchmarking of Simate‑beta on external robot testbeds to confirm the RoboDojo results; (2) announcements of model specifications, hardware requirements, and pricing that would indicate commercial readiness; (3) the rollout of the AutoResearch platform to external partners and any open‑source releases that could enable broader adoption; and (4) regulatory or safety assessments as the system moves toward more complex, real‑world tasks.
Independent third‑party evaluations of Simate‑beta on standard robot benchmarks (e.g., OpenAI Gym Robotics, Real‑World RL Suite) to confirm the leaderboard scores.
Public release of technical specifications, including model parameters, latency, and hardware platforms needed for deployment.
Announcements regarding the open‑source components of the AutoResearch platform, which could enable broader community adoption and scrutiny.
Regulatory scrutiny or safety certifications as the system moves toward more complex manipulation tasks in uncontrolled environments.