Simate, a recently formed robotics startup, says its AutoResearch system connects model modification, data and configuration changes, training, evaluation, and subsequent experiment decisions as part of its physical-AI research workflow, according to the 36Kr/Zhidx report. Reporting from 36Kr and Zhidx on September 24, 2026, describes a platform Simate calls AutoResearch that the company presented as an interface for moving from multiple experiments to a consolidated policy, while its official site identifies related branded components as Sinfra, Sipai, and RoboScientist.
The 36Kr/Zhidx piece attributes to Simate a workflow in which an agent can inspect model code and training configurations, adjust components through a pluggable model framework, invoke infrastructure to run training and evaluation, and then decide whether to continue, adjust, or stop experiments. Those operational details are presented as company disclosures in the 36Kr account and are not independently audited here. Simate’s public website uses the English product name Sipai for its model framework and shows RoboScientist as a research archive concept; the site also carries a ring-placement robot task illustration and a RoboDojo status line that indicates ongoing evaluation.
Early reporting and the company site can be read together alongside related industry materials. A 36Kr account of the company supplies the media narrative about AutoResearch. Simate’s official site describes its products and services for physical-AI work. The RoboDojo benchmark paper defines the evaluation framework Simate cited. Separate EastFrontier coverage of JD’s physical-AI work and XIRRA is related reading only and is not evidence for Simate’s claims.
The available account is a 36Kr report on Simate, supplemented by Simate’s official site and the RoboDojo benchmark paper. It arrives after JD’s earlier world-model release for physical AI and XIRRA’s recent physical-AI model financing, which illustrate the wider but varied physical-AI field.
AutoResearch and the company narrative
According to the 36Kr/Zhidx report, AutoResearch is intended to automate iterative research steps that typically require human orchestration. The report describes screens labeled From Experiments to One Policy and cross-scenario evaluation, and it characterizes AutoResearch as able to access historical experiment records and to coordinate training and evaluation infrastructure. The company told the reporter that AutoResearch is open for trial and that it has launched internal-test recruitment, claims that should be read as vendor disclosures rather than third-party verification.
Simate’s official English pages on its corporate site present a somewhat different naming structure. The homepage and an About page describe the firm’s mission with branded components Sinfra, Sipai, and RoboScientist and say Simate builds a platform, model, and an AI scientist to move physical intelligence from idea to action to insight. The Sipai research page marks training workflows and deployment as in development and coming soon, and the RoboScientist gallery lists a single saved study dated September 16, 2026 with 16 evolution nodes and 25 runs. That gallery entry illustrates an archive approach but does not establish that RoboScientist and the AutoResearch system described in the 36Kr report are identical products.
The 36Kr story attributes to Simate a taxonomy it calls Physical RSI, a proposed route linking model development, data processing, experiments, and real-robot deployment, with agents taking increasing responsibility for research tasks. The report divides RSI into weak, medium, and strong categories and places the company’s early work in the weak-to-medium spectrum; these categories are Simate’s conceptual framing as reported by 36Kr rather than an accepted field taxonomy.
RoboDojo is central to how Simate frames its progress. The company said a newly developed general-purpose physical fast system model ranked first on RoboDojo about three months after the startup’s founding. That ranking is a company disclosure in the 36Kr account and is not independently verified here. Simate’s official site itself cautions that RoboDojo evaluation is ongoing, which creates a material qualification for any claims about public leaderboard status.
The RoboDojo benchmark paper is public. The July 5, 2026 arXiv paper describing RoboDojo documents a unified sim-and-real benchmark for generalist manipulation composed of 42 simulation tasks, 18 real-world tasks, and an evaluation that integrated 30 policies. The paper sets the benchmark scope but does not mention Simate or corroborate the company’s claimed ranking.
Technical claims, demos, and limits
The 36Kr/Zhidx report attributes several technical claims and a real-robot demonstration to Simate. The story says Simate used AutoResearch in a human-machine co-driven mode to produce a general-purpose physical fast system model and that the company demonstrated continuous, multi-step tea making with a robot that retained and used historical observations. The report also attributes to Simate a 4D physical perception and memory mechanism that organizes depth, geometry, motion, and contact relationships around task progress and continuous-state tracking.
Simate’s pages separately describe work on efficient visual encoding, spatiotemporal modeling, and feature compression to enable larger models to operate in real time on edge devices, and they mention a future architecture pairing a physical fast system with a slower general reasoning system. The opened official pages, however, do not publish model sizes, parameter counts, architecture diagrams, benchmark numbers, or reproducibility artifacts, and the RoboScientist gallery shows research in development rather than finalized public results. The RoboDojo paper provides a benchmark framework and task counts but does not validate Simate’s claimed ranking or demonstration outcomes. All of these distinctions matter for readers evaluating the firm’s technical progress.
How this fits the physical-AI landscape
Simate’s framing of an automated research loop echoes broader industry conversations about closing the experiment-design cycle between simulation, training, and real-world testing, but the specific claims here are company-originated and require external verification. The RoboDojo paper supplies a public benchmark architecture and task counts that can serve as a point of comparison for anyone seeking to verify vendor disclosures. Simate’s website supplies product naming and early demonstration artifacts while marking many features as in development, and the 36Kr story supplies a narrative and company-reported milestones including a claimed RoboDojo ranking and a multi-step tea-making demo.
Readers evaluating Simate’s progress should consider the difference between company disclosures and independently audited benchmark results. Simate says it has raised cumulative financing in the hundreds of millions of RMB, although the stage and exact amount were not disclosed; it has also posted early research artifacts. RoboDojo’s arXiv paper defines the benchmark that Simate cited, and Simate’s site displays a research-archive entry and product names. Independent leaderboard confirmation, reproducible metrics, and published technical reports were not found in the materials opened for this story. Simate’s AutoResearch loop is therefore best understood as a vendor-presented research workflow and an active development program that claims early successes while still posting many features as in development.
Simate’s approach places an automated experiment-management loop at the center of physical-AI work. Whether that loop yields reproducible, benchmark-verified breakthroughs will depend on public results, independent leaderboard updates, and technical documentation that can be examined against RoboDojo’s task suite and evaluation protocol.