Anish Diwan

Robotics & Machine Learning Researcher at @ias_tudarmstadt

[Nico Bohlinger @NicoBohlinger]

@anishhdiwan added six Imitation Learning and Inverse RL algorithms to RL-X: GAIL, AIRL, AMP, NEAR, TRIRL (with or without differentiable trust region layers).
Implemented in Jax and PyTorch, and fully jitted for maximum speed 🚀
Feel free to check them out: https://t.co/6OEyLqlhWY

[Anish Diwan @anishhdiwan]

🔹Contribution: We show this reward update is equivalent to mirror descent.
🔹Contribution: We show that policy search over these rewards is approximate entropic mirror descent, with bounded errors in the trust-region setting. This complements our earlier work, Trust Region IRL.

[Anish Diwan @anishhdiwan]

📢 New @NeurIPSConf paper with @Jan_R_Peters & @OlegArenz :
"A Geometric Perspective on Reward Function Updates in Inverse Reinforcement Learning"

We show that natural-gradient reward updates in inverse RL require no Fisher matrix at all 🧵👇

[Anish Diwan @anishhdiwan]

🔹Contribution: With the Fisher information averaged along the straight line to this target, the natural gradient is just a convex interpolation of the current reward and the log occupancy ratio.

That ratio is cheap to learn with a classifier, so no Fisher matrix is needed.

[Aditya Bhatt 🤖🛹 ✈️ CoRL @aditya_bhatt]

A first glimpse of the wildest robotics project of my PhD.

🛹 The Blind Skateboarder 🛹

No vision. No markers.

Paper, code, weights — coming soon. And we have more. 😉

[Anish Diwan @anishhdiwan]

Thu, Jul 9 2:30 – 4:15 PM in Hall A (# 213)

Come by to discuss IRL, imitation learning, and how trust region policies are a cheap alternative to globally optimal max-ent IRL 😄

PS: I’m also co-organising the “Robot Learning Social” with @JoeMWatson on Wed 7-9 PM in Room 300.

[Anish Diwan @anishhdiwan]

There’s been a decade-long problem in inverse RL: learning an optimal reward function requires repeated RL solves in the inner loop. We show that this is no longer necessary!

I’m at @icmlconf presenting “Trust Region Inverse Reinforcement Learning” (https://t.co/GDFRtq8uFD)!

[Joe Watson @JoeMWatson]

@anishhdiwan and I are organising a social for robot learning researchers at #ICML2026, please swing by tomorrow evening to Room 300 if you want to chat 🤖🧠⚡️

[Kai Ploeger @KaiPloeger]

Learning five-ball juggling on the second attempt, with two Barrett WAMs. Most humans take years of practice.

#robotics #juggling

[Anish Diwan @anishhdiwan]

Huge thanks to my co-authors @davide_tateo, Chris Mower, @hbouammar, @Jan_R_Peters, @OlegArenz. And to @ias_tudarmstadt, Noah's Ark Lab, @Hessian_AI, @DFKI, @roboticsgermany for their support.

[Anish Diwan @anishhdiwan]

📢 Excited to share that our paper 𝗧𝗿𝘂𝘀𝘁 𝗥𝗲𝗴𝗶𝗼𝗻 𝗜𝗻𝘃𝗲𝗿𝘀𝗲 𝗥𝗲𝗶𝗻𝗳𝗼𝗿𝗰𝗲𝗺𝗲𝗻𝘁 𝗟𝗲𝗮𝗿𝗻𝗶𝗻𝗴 (𝗧𝗥𝗜𝗥𝗟) has been accepted at @icmlconf 2026!
🌐 Webpage: https://t.co/JCvZIq6NiJ
📄 Paper: https://t.co/GDFRtq7WQ5

#ICML #ReinforcementLearning #Robotics

[Anish Diwan @anishhdiwan]

🔥 We get desirable properties like monotonic improvement and a globally meaningful reward.
📊 2.4× improvement over SOTA IL baselines on MuJoCo & Robotics tasks (Unitree G1/Go2).
📊 Learns rewards that capture the expert's intrinsic motivation and transfer to new dynamics.