Writing · Pranay Gupta

03 writing

Notes on robot learning: making manipulation policies fast and general, simulation that holds up against real robots, and reinforcement learning after pre-training.

2026

Can frontier LLMs fix robots by editing their weights?Nov 4 We asked Opus to improve π0.5 by editing its weights directly, with no gradient training. Laundry folding went from 51% to 75%, and a mahjong task from 23% to 36%.
Discovery-0: zero-shot sim-to-realArgon Robotics ↗Oct 21 A model trained only on trajectories from our simulated kitchens closed a commercial dishwasher in a live customer kitchen, with no real-world data for the task.
The Turing test for virtual worldsArgon Robotics ↗Oct 14 Ten frames from the same robot stations, some from the real robot and some from our simulation of it. A game: can you tell which is which?
Scaling and speeding up robots in the real worldSep 18 Three levers that make robot policies run nearly four times faster than the remote-teleoperation data they learned from, without losing accuracy.