Manifold AI Tops RoboDojo, Beating GPT-6-Astra and GR00T-N1.7

Manifold AI has taken the top spot in the Agent Track of RoboDojo, a benchmark that evaluates robot learning. GPT-6-Astra and NVIDIA GR00T-N1.7, two prominent reference points for general-purpose robotics models, finished behind it. The win came from a system called R²SI. It grew out of work from three months ago, when Manifold AI released a world action model with In-Context Learning (ICL).

What we know about the result

The source reports three things. Manifold AI beat its competitors specifically in the agent track, not in the overall table. The models it passed include both a general-purpose language model at the level of GPT-6-Astra and NVIDIA's specialized robotics model GR00T-N1.7. The lead came three months after the release of the previous ICL model.

The available excerpt contains no numbers: no scores, no margin over second place, and no count of tasks in the benchmark. It also gives no expansion of the name R²SI. It is too early to judge the scale of the win. A narrow first place and a confident lead across all scenarios look identical in a leaderboard table.

Why ICL matters for robots

A world action model predicts how the world will change after a robot's action and selects the action itself at the same time. Add ICL, and the model can adapt to a new task from a few examples in its context, without retraining its weights. For a lab, that saves weeks of data collection and fine-tuning. For a customer, it shortens the path from "we want the robot to do it like this" to a working prototype.

The comparison with GPT-6-Astra is telling for a different reason. Large general-purpose models plan and reason well, but contact physics, latency, and execution errors remain a weak spot. If a specialized system beats them in an agent track, that supports the case for models trained on action data and world dynamics. GR00T-N1.7 belongs to exactly this class, so beating it is not easy either.

What is confirmed and what is not

• Benchmark: RoboDojo, Agent Track.
• Result: first place on the leaderboard.
• Beaten: GPT-6-Astra and NVIDIA GR00T-N1.7.
• Predecessor: a world action model with ICL, introduced three months ago.
• New system: R²SI.
• Not disclosed in the source excerpt: final scores, model size, training data, availability of weights and API.

What teams should watch

Until a technical report appears, it is reasonable to treat the result as a strong signal, but not as grounds for changing your stack. Four things need checking. First: whether Manifold AI publishes its evaluation protocol, and whether independent teams can reproduce the numbers. Second: how well the Agent Track gains carry over to your own tasks, such as manipulation, grasping, and handling unusual objects. Third: how many in-context examples R²SI needs to adapt, and how inference latency grows. Fourth: what kind of access will be offered: open weights, a cloud API, or only partner pilots.

For those building robots on GR00T or on top of general-purpose LLMs, a sensible step for now is this: when details are released, run your own scenarios on RoboDojo and compare the results with your pipeline on your own data. If Manifold AI's ICL works as claimed, the gain will be in how quickly new tasks can be deployed, not just a line in a leaderboard.