
I'm an undergrad at Stanford studying computer science and philosophy. I work on memory and continual learning for generalist robot policies at the Stanford AI Lab, where I'm part of the IRIS group advised by Chelsea Finn. I previously spent a summer at Google working on Gemini Enterprise for CX.
Long-horizon manipulation requires memory, yet symbolic subgoals and perceptual history each excel on different tasks. We introduce a hybrid-memory π0.5 policy that combines both: a fine-tuned VLM writes subgoals into the policy's language prompt, while a bank of sampled past frames conditions its action expert directly. On the 16-task RoboMME benchmark, we show that the hybrid reaches 52.3% success, outperforming symbolic-only (45.7%) and perceptual-only (46.8%) memory.
Humanoid VLAs trained by imitation cannot distinguish good demonstrations from bad ones. We present the first reproduction of Physical Intelligence's RECAP on an open-source model, applying advantage-conditioned fine-tuning to Ψ0 with only 500 simulated rollouts. On a whole-body bend-and-pick task, we show that advantage conditioning improves success from 18.2% to 30.8% over behavior cloning on the same data.
Last updated: October 2026
aadi2 [at] stanford [dot] edu