Anish Ravichandran

AI x physical world
MIT BS '24 MEng ‘25
Looking for roles in physical AI

[Anish Ravichandran @anishrav1]

Our winning setup paired our fine-tuned Dream-1 checkpoint with a VLM planner + CV-based state tracking, and it beat other teams in a surprise 1v1 playoff!

[Anish Ravichandran @anishrav1]

Second robotics hackathon, another first place!

Had a great time at the @makermodsai x @dreamlabs_ai hackathon this weekend with @joeeriqat, @n0homie, and @IsaiahBallah! More details in 🧵

[Anish Ravichandran @anishrav1]

We finetuned Dream-1 on a USB unplugging task using just 50 examples and 500 training steps, and it worked remarkably well on a new embodiment and environment

[Anish Ravichandran @anishrav1]

@4gatepylon Crazy work for a single weekend 🔥

[Anish Ravichandran @anishrav1]

Entered my first robot hackathon and won!

Big thanks to @viamrobotics and @MongoDB for sponsoring, @RoboticsCtr for hosting, and to my insanely cracked teammate @joeeriqat!

[Anish Ravichandran @anishrav1]

@joeeriqat @MongoDB @viamrobotics Couldn’t have asked for a better teammate 🦾

[Joe Eriqat @joeeriqat]

First time entering a hackathon, first time winning a hackathon.

Massive thank you to @MongoDB and @viamrobotics for hosting and especially @anishrav1 for being a goated teammate

[Anish Ravichandran @anishrav1]

Anyone going tomorrow and wanna team up?

[Robotics Center of Silicon Valley @RoboticsCtr]

Your code. Real robot arms. 48 hours to build.

Robot Hackathon with @MongoDB and @viamrobotics at Robotics Center, Oct 29–30.

Build physical AI applications using real hardware, Viam and MongoDB Atlas.

Event details: https://t.co/AqwM2CTgPI

#PhysicalAI #Robotics

[Anish Ravichandran @anishrav1]

3/ testing a hybrid policy with consistent handoff

on 10 selected sim starts: pure ACT made 2 loops. expert bracelet approach → ACT made 6 clean loops.

is it worth pursuing an end-to-end ACT policy or to keep working on the hybrid? the next task variant moves the tower too

[Anish Ravichandran @anishrav1]

teaching an SO-101 to loop a bracelet over a tiny Eiffel Tower with ACT + @LeRobotHF.

real-robot clip below. the hard part is generalizing to new starting positions.

how would you improve spatial generalization without just collecting more demos?

more in 🧵

[Anish Ravichandran @anishrav1]

2/ more coverage helped, but didn't solve it

271 sim episodes, 100k ACT updates on an L40S:
10/12 on training starts
7/20 on new spatial holdouts

seems like spatial interpolation needs more training data and/or closer starting position clusters