person 0.96
Pranay Gupta
01 about
We are digitizing the physical world. Everything that has been digitized can be learned.
At Argon Robotics we build simulated worlds where robots learn, and fail, before they ever touch the real one.
We’re contributors to Newton and MuJoCo Warp, the physics engines NVIDIA and DeepMind build on, and to open research like Terminal-Bench.
We’ll be at CoRL, and we’re hiring. Come say hi, or email me if you’re up for a good conversation about robots, world models or LLM research.
education
- B.Tech, Computer ScienceIIT Delhi · 2019–2023All India Rank 82, JEE Advanced
Interested in knowing what I fix?
- newton#4791 sphere and capsule radius after GJK/MPR
- newton#4795 direct solve for stiff, slender rods
- mujoco_warp#1796 EPA sliver faces
- mujoco_warp#1801 float64 Hessian for stiff contacts
- terminal-bench#2492 distributed dedup run time
- newton#5146 deterministic VBD attachment forces
- newton#5137 MJCF noslip options in SolverMuJoCo
- newton#5131 scale-free VBD bending test
- newton#5128 one-value MJCF solref padding
- newton#5121 MuJoCo pair properties after reordering
- newton#4977 changelog for barrel-cylinder contacts
- mujoco_warp#1802 constraint-free worlds in sparse solve
02 featured
new 0.93
Can frontier LLMs fix robots by editing their weights?
Claude edited π0.5's weights with no training: laundry folding 51% → 75%, a mahjong task 23% → 36%.
Nov 4, 2026 · blog post
demo 0.87
Inside a robot's brain
All 73,728 of π0.5's motor neurons, firing live as the robot folds a shirt. Tap any one to see what it does.
Nov 2026 · interactive ↗
result 0.91
Discovery-0: zero-shot sim-to-real
A model trained only in our simulated kitchens closed a commercial dishwasher in a live customer kitchen, with no real-world data for the task.
Oct 21, 2026 · Argon Robotics ↗
game 0.78
The Turing test for virtual worlds
Ten frames from the same robot stations, some real and some from our simulation. Can you tell which is which?
Oct 14, 2026 · Argon Robotics ↗
result 0.84
Policies faster than their demonstrations
π0.5 on two PiPER arms does a bimanual dish task in 9.4 s, against 37.3 s for teleoperation, at 95.2% success over 500 real runs.
Sep 18, 2026 · Argon Robotics ↗
