Aadi Shah

Robots doing long tasks need memory, but language subgoals and raw visual history each win on different tasks. We train one π_0.5 policy that uses both, reaching 52.3% success on the 16-task RoboMME benchmark versus 45.7% and 46.8% for either alone. A fine-tuned VLM writes subgoals into the policy's language prompt, while a bank of sampled past frames steers its action expert directly.

Humanoid VLAs learn by imitation, so they can't tell their good attempts from their bad ones. We ran the first reproduction of Physical Intelligence's RECAP on an open-source model, Ψ₀, using just 500 simulated rollouts. Advantage conditioning beat plain behavior cloning on the same data (30.8% vs. 18.2% success) on a whole-body bend-and-pick task.

Testing small drones in simulation is still mostly manual, which limits how many scenarios developers can cover. We propose AutoSimTest, where LLM agents collaborate to write test scenarios, configure the simulator, generate missions and analyze the flight logs. On 10 real PX4 incidents, its diagnoses scored 9.5/10, closely matching those of experienced PX4 developers.