On September 23, Simate-beta—a robotics model from a Chinese startup barely three months old—claimed the top spot on the RoboDojo Sim leaderboard, posting an average score of 33.95 across 42 manipulation tasks. Five days later, it had slipped to second place.
The volatility captures the pace of competition in embodied AI. RoboDojo, a benchmark launched in July 2026 with cloud-based anti-cheating verification, has attracted more than 30 models from labs spanning Xiaomi Robotics, Google DeepMind spinouts, and university teams across China, the U.S., and Hong Kong. By September 28, "PhysicalRSI Agent + VLA" from HKU MMLab and KAI held first place with a 36.27 score and 31.38% success rate, according to the live leaderboard maintained by XPolicyLab.
Simate, the startup behind the displaced model, is using what it calls an "AI scientist" to automate research cycles from hypothesis to experiment to feedback. If the approach works at scale, it could compress development timelines in a field where simulation performance doesn't always translate to real-world deployment.
The company positions itself as "AI FOR PHYSICAL AI," according to its homepage, and operates three products: Sinfra (infrastructure), Sipai (the model), and RoboScientist (the automation layer). As of early October, its site listed RoboDojo evaluation for Sipai as "in progress."
Fangneng Zhan, an assistant professor at Hong Kong University of Science and Technology who leads the World Mind Lab, serves as Simate's chief scientist. Zhan wrote on his personal site in September that "our Simate team topped the RoboDojo Leaderboard." Zhang Ying, the company's founder and CEO, was described by the company in Chinese tech coverage as a "former core tech lead" on a major autonomous-driving program in China, though the specific prior employer was not named in available sources and the claim has not been independently verified.
The Broader Automation Shift
Simate's research automation strategy mirrors trends elsewhere. Anthropic said in September that Claude "leads" 26% of its AI R&D work as of August 2026, up from less than 1% in February, according to the company's R&D Automation Index. Microsoft Research published a conceptual paper in May titled "AutoResearch AI: Towards AI-Powered Research Automation for Scientific Discovery."
Multiple Chinese outlets reported in late September that Simate "completed multiple rounds of RMB hundreds-of-millions financing" within three months of founding. No investor names, amounts, or round types appeared in filings or primary documents as of early October, and the claims remain unverified by independent sources. The company declined to disclose headcount, named customers, or specific use cases for its platform.
RoboDojo tracks manipulation skills across five dimensions: generalization, memory, precision, long-horizon planning, and open-ended instruction following, according to the benchmark's paper published in July. The simulation leaderboard runs continuously. A separate real-world board, which uses a standardized robot setup called RoboDojo-RealEval for 18 physical tasks, showed OpenWAM-α leading as of September 28 with a score of 37.60 and a 24.40% success rate.
The benchmark includes models from Xiaomi Robotics, Google's InternVLA_A1, OpenVLA-OFT, Physical Intelligence's Pi-0 and Pi-05, and systems from startups like StarVLA, GalaxeaVLA, and EventVLA. On the real-world board that day, Pi-05 ranked second with a score of 22.90 and a 12.80% success rate; InternVLA-A1 placed third at 12.00 and 7.20%.
Infrastructure and Policy Context
The benchmark surge is unfolding as robotics funding and policy support accelerate. PitchBook data showed $27.6 billion in robotics funding across 1,009 deals in 2025, according to a summary published in May 2026. Agility Robotics, the humanoid maker, went public via SPAC in June 2026 at roughly $2.5 billion, according to SEC filings.
China's Ministry of Industry and Information Technology issued "Humanoid Robot Innovation Development Guidance" in November 2023 and a "Robotics+ Application Action Plan" in January 2023, setting national priorities and 2025 targets. Local governments followed. Hangzhou passed regulations in December 2025 to promote the embodied-intelligence robot industry. Shanghai's 2025 plan prioritized embodied AI research. China's standards registry in June 2025 added "Artificial intelligence—Technical requirements for embodied large model system" as a guidance document project.
The International Federation of Robotics said China installed 295,000 industrial robots in 2024—54% of the global total—in a press release published September 25, 2025. IFR President Takayuki Ito wrote that "comparison with China reveals the enormous automation potential" for other markets, including the U.S., which runs the third-largest robot stock worldwide at 393,700 units.
ABI Research estimated China accounts for 97% of current humanoid deployments, projecting that share to drop to 72% by 2030. Research & Markets forecast the commercial humanoid market at roughly $900 million in 2025, growing to around $7 billion by 2030. Grand View Research pegged the broader embodied AI market at $4.67 billion in 2025, projecting $67.63 billion by 2033 at a 39.7% compound annual growth rate. Morgan Stanley published a note in April 2025 projecting a potential $5 trillion humanoid market by 2050, with China holding the largest installed base.
Benchmark Credibility Questions
A paper published in June 2026 titled "What Are We Actually Benchmarking in Robot Manipulation?" warned that cross-benchmark inconsistencies and overfitting risks complicate ranking comparisons. The authors advocated for diagnostic checks across multiple suites.
Community variants like Robocurve's "RoboDojo-RC Tier 1" adaptation, published September 23, use different embodiments and custom rubrics, making them not directly comparable to the official board. GPT-6 Astra, an LLM-as-policy system, claimed a 28.97 score and 22.48% success rate over 2,100 simulation trials in a preprint posted in September, but the work is not an official RoboDojo entry. The preprint also reported safety stops during real-world trials, and the study halted its real-robot protocol.
RoboDojo's governance model features continuous leaderboard updates with anti-cheating verification, as described in the benchmark's GitHub README. That contrasts with static academic benchmarks and may pressure teams to submit models earlier in development.
What This Means for the Field
AI-automated research systems entering robotics labs could shorten iteration cycles between simulation and physical deployment, though the gap between benchmark performance and real-world readiness remains substantial. The U.S. has measurement-science programs tracking the trend: NIST's 2026 program documents reference next-generation robotics goals and physical AI data generation for manufacturing.
China's combination of industrial robot dominance, national robotics policy, and embodied AI standards development creates structural support for rapid commercialization pathways that differ from U.S. patterns. The policy environment, coupled with the automation share advantage, suggests that the benchmark competition will intensify.
Expect more startups to run "auto-research" pipelines in the months ahead. Expect more attempts at LLM-based policies with stricter real-world safety gates. And expect continued jockeying around standardized sim-to-real evaluation protocols like RoboDojo-RealEval and the BEHAVIOR challenge, based on trends visible in Anthropic's R&D automation data and the benchmark landscape as of late 2026.
Simate's quick ascent sits inside this acceleration. It also sits inside the lingering question of whether benchmark rankings predict deployment readiness or just simulation performance. The five-day rankings shift in September offers a partial answer: the field is moving faster than any single snapshot can capture.
