Srinath Mahankali

After an amazing few years at MIT working on RL algorithms and robot learning under the incredible mentorship of @pulkitology, following Pulkit to @EkaRobotics was a no-brainer. Seeing the raw speed of the robots in person, I knew immediately that I had to be part of the mission. It was a total blast in those early months building alongside Pulkit, @haarnoja, and @kristianhartika, and it’s been amazing to work with all the brilliant people who have joined the team since. It's been such a great opportunity to build on the speed of our robots and drive our generalization capabilities forward—by figuring out how to get our systems to use vision, tackle highly cluttered scenes, and yes, screw in lightbulbs 💡 Eka is proving that robots can be general, fast, and dexterous all at once, making it a genuinely special place to build the future of robotics—and we are hiring!

@skyfallai Never being continual learners feels a bit strong and it probably also depends on what tools they have access to. Were they able to spin up subagents, modify their own context, or finetune themselves for instance?

Today we present Morpheus, a persistent enterprise simulation platform designed to make Continual Learning a reality. Morpheus is the world’s first real world Reinforcement Learning environment. Every Reinforcement Learning environment operates in the game world. Benchmarks like Atari, OpenAI Gym, MuJoCo, and Procgen are all small, game-like worlds that reset every few minutes. But the real world never resets. A business keeps running and evolving everyday. We tested how frontier LLMs would perform in realistic and dynamic business environments 🧬on Morpheus. The main conclusion was that LLMs are not continual learners. 🧵Here’s how we did it and what we learned:

@fujikanaeda if theres more reward to being correct it sounds well-specified, unless the episode ends upon correctness

reward hack: if the env provides a reward when taking correction from the user, then the hack is to sandbag so you can get the reward when they correct you.

@alex_raistrick This is very cool work!

1/ We've released Infinigen 2.0! Currently in preview. It creates indoor 3D scene files in 1min CPU time, and includes new and better materials --- all still fully procedural. Our new 2.0 design is highly efficient and allows easy control and recombination via Python APIs.

@gabriel1 they could have ndas about the projects though

the reason why it's hard to talk about is that no one expects it to come up, and you actually need to think really hard to come up with great examples of deep insight i used to think hard 3 hours before every interview to optimize how i communicate

after now having interviewed 50+ people, it's crazy how rare it is to talk about things you've done in concrete terms always only talk with high precision about what you did and why no one else could it doesn't mean anything that you "made X better" or "led thing Y" examples:

@Hongyu_Lii @BrownCSDept @ColumbiaCompSci @MIT @Columbia @BrownGradSchool @FuWanjia @xiaoyan_cong @kunkun0w0 @binghao_huang @jiang_hanxiao Agreed! its time for agentic world modeling

@srinathm1359 @BrownCSDept @ColumbiaCompSci @MIT @Columbia @BrownGradSchool @FuWanjia @xiaoyan_cong @kunkun0w0 @binghao_huang @jiang_hanxiao We might be able to use some agentic systems to auto collect such data 🤔

Can a world model 🌏 learn to predict how 198 different deformable objects move -- not just one rope or cloth? That question motivated Deform360: 1,980 real-world interactions captured from 41 synchronized views with bimanual touch sensors, which becomes crucial when vision is occluded. Accepted at #ECCV2026. 🧵 1/5

Really cool work on modeling deformable object dynamics!

@Hongyu_Lii @BrownCSDept @ColumbiaCompSci @MIT @Columbia @BrownGradSchool @FuWanjia @xiaoyan_cong @kunkun0w0 @binghao_huang @jiang_hanxiao This is extremely cool!! I wonder if you could autonomously collect such interaction data if we wanted to scale this up.

This is a collaboration between @BrownCSDept, @ColumbiaCompSci, and @MIT. My visit to @Columbia was funded by @BrownGradSchool Research Mobility Fellowship program. Thanks for my collaborators @FuWanjia @xiaoyan_cong @kunkun0w0 @binghao_huang @jiang_hanxiao Xintong He @YiqingLiang2 @RaoFu79761158 @isaac__lt @drsrinathsridha Kevin A. Smith, George Konidaris @BrownBigAI , and @YunzhuLiYZ.

How can generalist policies adapt to new challenges at deployment using skills they already have? We optimize VLA *prompt inputs* with reinforcement learning, enabling efficient real-robot adaptation on complex tasks where existing methods struggle. 🧵 semantic-action-rl.github.io

Really cool work bridging sim2real gap!

@tylerlum23 @kushalk_ @leto__jean Amazing work!! so nice that the sim to real is zero shot :)

To support further research in dexterous assembly, we are releasing our code, assets, and training/evaluation environments. Full details + videos + code: play2perfect.github.io This work wouldn't be possible without @kushalk_ (co-lead), @leto__jean, and Prof. Karen Liu. 🙏

🤖 How can we teach dexterous robots to perform precise, contact-rich assembly? Introducing Play2Perfect: first learn to play with objects, then perfect the policy for tight insertion, multi-part assembly, and screwing. Sound on! 🔊 🧵👇

@fanshi_robot Super cool work!!

MPPI, teacher-student learning, and VLA — all with 100% simulation data, easy, cheap, scalable. 🚀 Deformable–rigid interaction has long been a bottleneck for fast & accurate simulators. Our GPU-native synthetic data pipeline breaks it for deformable sim2real. 🧵

@junyi42 @letian_fu This is very cool!!

𝐑𝐀𝐓𝐬 is a first step toward 𝐏𝐥𝐚𝐲𝐟𝐮𝐥 𝐀𝐠𝐞𝐧𝐭𝐢𝐜 𝐑𝐨𝐛𝐨𝐭 𝐋𝐞𝐚𝐫𝐧𝐢𝐧𝐠: 🌐playful-rats.github.io We see a future where the next step for agentic robots isn't just stronger test-time harness, but a play stage where they set their own goals, fail, and build up skills long before we hand them a task. Huge thanks to the team: @lukehanjun (co-first) @letian_fu, Zihan Yang, Yaowei Liu, Raj Saravanan (core contributors), @istoica05 @akanazawa @JiahuiLei1998 @HavenFeng @trevordarrell and many others!

Children learn from play. Can robots do the same? We propose 𝐏𝐥𝐚𝐲𝐟𝐮𝐥 𝐀𝐠𝐞𝐧𝐭𝐢𝐜 𝐑𝐨𝐛𝐨𝐭 𝐋𝐞𝐚𝐫𝐧𝐢𝐧𝐠, a paradigm that gives embodied coding agents a play stage before downstream tasks arrive, and instantiate it with 𝐑𝐀𝐓𝐬 (Robotics Agent Teams), where robots discover reusable skills through curious play. Co-led with @jiaxin_ge_

@york1to @yacineMTB this is insane

You can use mujoco warp to compile your simulation as a cuda graph, and then launch that cuda graph from C to have absurdly fast robotics simulations. Nvidia warp also open sources a bunch of really great renderers, fluid sim stuff

@yacineMTB i bet stacking a couple of timesteps would also work, it should be pretty close to fully observable. you might be able to get rid of the episode length randomization that way too if you want

@srinathm1359 gp says yea

behold. THE WORLDS FIRST SIX PENDULUM CARTPOLE SOLVE. Including a sponsor! To solve this task, I built an environment to train an AI. This is what mechanize does, but for larger AIs. Apply! Salaries are up on their page Thank you to mechanize for sponsoring!

@yacineMTB do you even need an rnn for this?

So once the model started scoring well enough that it learned the whipping behavior, but struggled to keep it up longer than 10 seconds, I increased and randomized the episode length per episode Just a dumb trick I found experimentally to make these little rnns behave better

@yacineMTB this is really cool!!

New startup out of stealth in Cambridge MA: Eka Robotics. The company is building intelligence for the physical world in its native language: FORCE. Their core solution to mastering fast, reliable adaptive robots: Vision-Force-Action (VFA) model: - sim-only reinforcement learning - no human teleop data - robot practices thousands of hours in a physics-rich simulator (mass, inertia) and comes up with its own solutions, AlphaZero-style. - custom grippers add touch sensing; the model maps pixels + felt force to actions.

@StephanJaeckel @TheHumanoidHub @EkaRobotics Everything is 1x speed! Except for the last one - we slowed it down to 1/25x :)

@TheHumanoidHub I stopped trusting videos of robots which do not show the speed at which they run. Sorry @EkaRobotics for the mistrust. But congrats if the vids are not speeded-up.