MIT '27 — B.S. Computer Science & Engineering, B.S. Management, MIT '28 M.Eng · Cambridge, MA

Shrish Choudhary

I am interested in teaching robots to learn autonomously through VLA manipulation, whole-body RL, hybrid RL/IL methods, and agentic systems.

Currently researching VLAs at the MIT Improbable AI Lab. Previously robotics at Multiply Labs and ML intern at eBay. Coauthor in Science Advances, IROS, and CLEO; co‑inventor on a robotics patent.

Email|Résumé|GitHub|LinkedIn

Shrish Choudhary waving beside Frank, a Unitree G1 humanoid robot, at Multiply Labs
Suspended mirrored spheres filled with light at teamLab Borderless, a photographer reflected in the nearest one
A band of orange light over a sea of clouds seen from high on a volcano in Maui
Close detail of the carbon-fibre rear diffuser and taillight of a white Koenigsegg Jesko
Shrish Choudhary photographed in the reflection of a round orange convex traffic mirror on a Kyoto street
Shrish Choudhary holding a camera, lit by a projected floral installation at teamLab Borderless in Tokyo
The undulating glass facade and cone-shaped mezzanine inside the National Art Center in Tokyo
A wave breaking over black lava rock in front of palms and Norfolk pines on the Maui coast
A bright green Bugatti Chiron Pur Sport parked on a street in Carmel-by-the-Sea
A colour-shifting McLaren P1 parked beside a light blue Ferrari 296 on a tree-lined street
A nutria peering out from dense green riverside grass beside still water
A red Lamborghini Temerario on gold wheels rolling down a street past onlookers
A green Lamborghini Revuelto seen from behind on a hillside street above Monterey Bay
1 / 16

Me and some photos/videos

Me with our G1 (Frank) in a One Piece Jersey @ Multiply Labs

Robots, and everything else — shot across Japan, Hawai’i, Monterey Car Week and MIT tennis on the road. Stills and short edits.

Click the frame to advance →

Selected work

Humanoid learning & autonomy · code and clips are public
Unitree G1-D · digital twin

G1-D Autonomy Stack

SLAM navigation, A* planning and a GR00T flow-matching VLA composed into one loop: navigate to the table, bimanual pick, carry across the room, place — behind a human accept gate and a LiDAR failsafe, validated in a MuJoCo twin before touching hardware.

GR00T · VLASLAM + A*MuJoCo twin

GitHub·Demo on LinkedIn

29-DoF G1 · force-reactive walk

Force-Reactive Locomotion

A 29-DoF G1 walks where it is commanded while its arms react to forces it has no sensor for. A small MLP infers per-hand wrench from proprioception alone; splitting the regression into separate direction and magnitude heads roughly tripled magnitude R². One policy covers carrying and pushing.

PPOForce estimationmjlab

GitHub·Results

Pedipulation · swing-foot tracking

Single-Leg Pedipulation

Balancing on one leg with the arms frozen around a payload, pushing an obstacle clear with the free foot. The actor sees only on-robot-computable observations, enforced by a deployability gate — 0.993 deterministic survival with zero torque saturation.

PPODeployable obsSim-to-real

GitHub·Results

Recognized work

Fine-tuning NVIDIA’s GR00T N1.7 to pick and place bioreactors on a Unitree G1-D at Multiply Labs, wired into deterministic navigation, LiDAR and collision avoidance.

The demo, posted by Multiply Labs’ CEO Fred Parietti of my work.

View on LinkedIn ↗

Experience

Publications & patents

Autonomous optical assembly at MIT RLE · precision to 0.1 mm / 0.1°

Also

  • AUTO‑Arm — autonomous optical-bench assembly, ~90% faster than by hand
  • TactileRL — contact-rich manipulation RL in Isaac Lab
  • Reflex — agentic RL/IL training pipeline
  • Skeye — on-device 360° robot safety perception
  • Mirim — 2D listings → 3D Gaussian-splat assets, built at YC IAP
  • Wound Screener — the Harvard Medical deployment, on the Play Store
  • YC IAP recipient
  • Harvard Medical School NTAC prize contributor
  • Eagle Scout
  • Scholar Athlete Award
  • Varsity tennis @ MIT