What Simate-beta’s RoboDojo Result Says About Its Physical AI Ambitions

Simate-beta is Simate’s first general-purpose physical-AI “fast system”—a robot-control model designed for quick action in the physical world. The RoboDojo leaderboard lists a Simate-beta entry with a score of 33.95 and a 27.96% success rate on its simulation leaderboard. Simate also describes separate real-robot demonstrations, but neither the leaderboard result nor selected demos establish how reliably the system will perform across everyday tasks and environments.

What the RoboDojo score does—and doesn’t—show

RoboDojo is described as a benchmark for evaluating generalist robot manipulation policies across simulation and real-world settings. The specific Simate-beta result listed on the leaderboard is a simulation entry, contributed by Simate—not a reported score from real-robot testing.

That distinction matters: the score is a benchmark result, while the company’s demonstrations are a separate way of showing the system operating on physical robots. Coverage of the demonstrations describes memory, task adaptation, long sequences of actions and fine manipulation. These examples indicate what Simate is trying to achieve, but they do not by themselves establish general performance beyond the demonstrated tasks. 5

A fast system built for responsive control

Simate describes its design as a fast system that combines 4D physical perception with memory. The company’s stated goal is to support responsive, potentially on-device control and eventually pair that fast system with slower reasoning for more complex tasks. 15

The idea is to track how a physical scene changes over time and retain useful information about task progress, while still reacting to new conditions. That design ambition connects the RoboDojo result and the physical demonstrations: both are presented as early evidence for a system intended to perceive, remember and act. They are not, on their own, evidence that the intended capabilities will generalize to unfamiliar situations. 10

AutoResearch: using agents to help develop robot models

Simate’s broader approach, which it calls “Physical RSI,” aims to connect model development, data work, experiments, evaluation and real-robot deployment in a recurring research loop. AutoResearch is intended to let AI agents assist with proposing and testing research ideas, while people provide goals and constraints and assess the results. Simate presents beta as an early output of this human-guided process. 1112

The company also says its in-house infrastructure connects training, inference and evaluation, and supports parallel research directions. Its Sinfra offering is described as covering simulation as well as training and inference. The practical claim is that a connected research setup could help the team test ideas and incorporate results more quickly; the available reporting does not independently establish how much it improves model performance. 1115

Team experience and planned releases

Simate’s reported team brings together autonomous-driving deployment experience and research backgrounds in areas including world models and real-robot control. That combination is relevant to the company’s stated goal of developing models that can both learn about the physical world and operate on hardware, though team credentials are not a substitute for independent performance evidence. 512

Simate says it plans to publish a staged result on zero-shot generalization to complex tasks by the end of 2026, followed by papers or technical reports on its model and automated-research work, with open-source releases phased over time. These are announced plans, not results or releases already available. 11

For now, the clearest read is a promising but company-led early report: a listed simulation score, separate real-robot demonstrations and a research pipeline designed to speed iteration. The next important evidence will be detailed technical reporting and results that clarify how performance holds up beyond the benchmark entry and showcased tasks.