I think distilling the model’s spatial knowledge into a raw action space policy would be an interesting direction. its good vision understanding makes this verifiable enough to happen in an automated closed-loop way

It used fewer reasoning tokens than other VLM APIs (though the counts might not be directly comparable). I couldn’t run the full benchmark for these API models to include them on our leaderboard (molmospaces.allen.ai/leaderb…) because of the API cost :)

All models used the same closed-loop harness in @robocurve’s results. they get exocentric/egocentric rgb observations from the DROID setup and robot state, and output cartesian poses. history is cleared after each of the 200 episodes.

GPT-Astra outperformed all open-source VLA/WAM baselines on a subset of our zero-shot MolmoSpaces-v1 benchmark! traces: orayyan.com/ms-results

Training for the robot olympics with MuJoCo’s new surfacevel for geoms

Check out our website for more videos and interactive simulation episodes: orayyan.com/fetchman. This project was done with advice from @YuchenCui1 and great collaborators @max_argus, Zhi Li, Chang Yu, Yuxin @chenfanfujiang. Paper: arxiv.org/abs/2608.17027 Code (soon): github.com/omarrayyann/Fetch…

We follow this sim-to-real recipe to train two policies: a single-object Fetch policy that works across scenes, and an object-conditioned policy that takes the target object's name as an instruction. Both are two-system architectures with the decoupled SONIC controller acting as the low-level controller.

Alongside FetchMan, we're releasing FetchMan-Bench, a simulation benchmark for humanoid policies that expands MolmoSpaces-Benchmarks toward humanoids, starting with the Fetch task. Thousands of evaluation rollouts can be run across scenes to give a quick, reliable read on how a humanoid policy is doing.

Post-training FetchMan with large-scale Flow-GRPO in sim lets the policy break beyond the confines of its synthetic demonstrations, using only a simple sparse reward on our Fetch task. We turned to RL because performance saturates at a ceiling that more of the same synthetic simulation data can't push past.