AgiBot's GO-3 Nano Preview Tops RoboDojo Simulation Leaderboard With 61.90 Average Score
A preview of AgiBot's next robot manipulation foundation model, GO-3 Nano Preview, ranks first on RoboDojo's simulation leaderboard with a 61.90 average score and a 56.42 percent average success rate. It leads three of six dimensions and trails other entries on memory and generalisation.
A preview version of AgiBot's next robot manipulation foundation model, GO-3 Nano Preview, has taken first place on the RoboDojo simulation leaderboard with an average score of 61.90 and an average success rate of 56.42 percent. RoboDojo is a public benchmark for generalist robot manipulation policies that combines 42 simulation tasks with 18 real-world ones. The details below come from reporting by Pandaily.
The numbers
RoboDojo's update log records the entry as added on October 27 and contributed by AgiBot, on the agent track. Second place is PyRUA-Lean from the PKU DAGroup team, at 55.26 average score and 50.35 percent success rate. The margin at the top is 6.64 points.
RoboDojo splits the simulation score into six dimensions. GO-3 Nano Preview does not lead all of them:
- Precision manipulation: 78.65, first place.
- Long-horizon execution: 68.21, first place.
- Open-ended semantic tasks: 61.71, first place.
- Memory: 64.50, behind HexaShell-Preview (84.53) and EmbodiedEvo (79.40).
- Standard generalisation: 45.93, behind RobotEra's VPP2 (52.14).
- Randomised generalisation: 26.97, well behind entries that specialise in this dimension.
Pandaily notes that Chinese coverage of the result quoted the same headline numbers, with some outlets presenting the per-dimension scores as uniformly first. The leaderboard itself is the stricter record, and the figures above follow it.
Classification and field
The benchmark classifies the model as an agent plus world-action-model system rather than a plain vision-language-action policy. Pandaily says this fits the direction AgiBot has taken since GO-1 and its GE-Act world-action work.
The simulation board lists dozens of policies from Chinese and international teams, including Shanghai AI Lab, RobotEra, Galaxea AI, Horizon Robotics, Tencent Robotics X, Li Auto and Xiaomi Robotics, alongside reference entries run by the benchmark team.
Separately, a system called RoboDoctor, from a team led by a Tsinghua University professor, reported 57.10 on RoboDojo's 54-task simulation summary this week.
Status of the model
AgiBot has not published a technical report for GO-3 Nano Preview. The leaderboard entry is marked as a preview, so the figures describe a work in progress rather than a shipped model. Parameter count, context window and pricing are not stated in the reporting.
What this means
The 61.90 average places GO-3 Nano Preview first overall, but the per-dimension breakdown shows an uneven profile. The model is strongest on precision manipulation, long-horizon execution and open-ended semantic tasks. It is weaker on memory and on both generalisation measures, where its 26.97 on randomised generalisation is the lowest of its six scores.
The overall rank therefore does not mean the model leads in every capability, and readers comparing policies for a specific use case should check the relevant dimension rather than the average. Because the entry is a preview with no technical report, the scores cannot yet be tied to a described architecture or training recipe, and they may change in a final release.
The result also shows AgiBot's model line, not only its robot hardware, being measured against other teams' policies on a shared public board.
Related Articles
ByteDance's Seedance 2.0 Pro tops HKU video AI ranking with 3.84 out of 5 for visual quality
A University of Hong Kong study ranked 12 leading AI video-generation models and found Chinese models ahead of US rivals on visual quality. ByteDance's Seedance 2.0 Pro scored 3.84 out of 5, while several Chinese models fall behind on safety safeguards, according to the researchers.
Swapping coding-agent harnesses moves SWE-bench scores about as much as a rerun, but costs can triple, study finds
Two Shandong University researchers held the model constant and swapped coding-agent harnesses. Pass rates on SWE-bench Verified changed about as much as a plain rerun, while per-task cost varied by up to three times. The paper is titled "What does a harness buy? Mostly tokens."
Guide benchmarks Cactus Whistle, a 16.9 MB speech model, against Whisper base on a CPU
MarkTechPost published a coding guide that benchmarks Cactus Compute's 16.9 MB speech model, Cactus Whistle, against Whisper base on a CPU. The guide covers word error rate, speed, streaming latency, timestamps, keywords and speech-to-tool calls.
Microsoft's MAI-DxO scores 83% vs 36% for physicians on complex NEJM cases in Science study
Microsoft's MAI-DxO, a multi-agent diagnostic orchestrator, was correct on 83% of complex New England Journal of Medicine cases in simulation, versus 36% for generalist physicians, according to a study in Science. The data suggests most of the gain comes from stronger base models, with orchestration adding a smaller lift, mainly on complex cases.
Comments
Loading...