Sattyam Jain

Async inference is the part I keep pointing people to. Has anyone on the team looked at how SmolVLA follows instructions phrased unlike the training data? In LIBERO I saw ordinary paraphrases do almost nothing while an odd framing did a lot, and I'd like to compare notes with whoever owns evals.

Good to see on-device numbers for VLA fine-tunes. After the on-device optimizations, did you compare the optimized policy with the original on the same episodes, including the failed ones? A port can keep its success rate and still move somewhere new when it fails, and on an embedded robot that's the part that matters.

EnvHub for evaluation makes a lot of sense. Would it also be a good place for shared perturbation sets, like a fixed list of reworded instructions per task with the original as the control? Right now everyone writes their own, so wording results from different groups can't be compared.