Aditya Bhatt

Introducing CrossQ, just published at #ICLR2024! 🎉 CrossQ achieves: 🔥 Very fast off-policy Deep RL 📈 with SOTA sample-efficiency in <5% of the gradient steps 🧹 without target nets or Q ensembles 🧑🏻‍💻Project and Code: adityab.github.io/CrossQ Joint work w/ @DPalenicek 🧵👇

It evolves again. 🦖 65 million years after they vanished, something awakens again—this time in a lab. DOBOT introduces FIRST BREATH Biomimetic Embodied Intelligence 3.0, our latest breakthrough in lifelike robotics. 🤖 #DOBOT #INFFNI #EmbodiedAI #BiomimeticRobotics

Really cool work. CrossQ's cross-batch BatchNorm trick strikes again!

@REK REK TEK is top-tier naming, keep it coming 😅

What's REK doing?!? It's not a whacky European Olympic sport, it's the (very scientific) reliability test of the REK TEK, to test whether our system can maintain control of our robot after being relocated 50+ feet. The experiment was a success😎

@ChongZitaZhang I guess any quadruped is a humanoid with the right *attitude* 📐

youtu.be/IiVReMBHpyI I thought anymal was not bioinspired, but it turns out a human morph...

xTerra is one of a small handful of companies making Quasi Direct-Drive motors (and robots) in India.

Need RL envs / simulation suite for Mining-adjacent tasks. call it “Minecraft”

At this point they can straight up do PPO on 4096 real robots and bypass sim2real entirely 😅 Only half kidding

Drop 12/14: Models, products, impact - today something different, very different. Launching Sarvam Kaze, our foray into getting our models into the your hands with our devices - designed and built here in India!

Insanely good

@alexanderawolf Yes! I’m working on this. It is possible. Paper coming up real soon 😉

Today I decided to close my eyes. I made my breakfast with my eyes closed. I washed the dishes with my eyes closed. I assembled a robot with my eyes closed. You get the point. And somewhere between the Colombian coffee, the Canadian maple syrup, the European neuromorphic chips, the Korean robotic hand, and the Chinese servos, it became clear to me: VLAs are not the way. Vision is the most computationally expensive signal the brain processes, and yet it’s only one of many; we can still function when we close our eyes. American robots should aim at least for that.

Huge thanks to my co-authors: @firasalhafez, @aditya_bhatt, @Jan_R_Peters, and @OlegArenz. Grateful for the support from TU Darmstadt, DFG, and Hessian AI. If you'll be at ICLR (Rio 🇧🇷), come say hi! Questions & feedback welcome 🙏 @ias_tudarmstadt #ICLR2026

From discrete latents to physical motion. DAPS serves as a compact command space: language + motion goals → discrete codes → decoded into physically consistent control. Simulated with LocoMujoco.

DAPS: Discrete Variational Autoencoding via Policy Search (accepted @iclr_conf) A principled way to train discrete autoregressive encoders — without straight-through gradients. Entropy reg. + KL trust region for stable training and compact latents. 🔗 drolet.io/daps

The future is software writing its own software. Which is why I'm so in love with Pi: a coding agent that can extend itself :) lucumr.pocoo.org/2026/1/31/pi/

@Ed__Johns Congratulations!

I'm very excited to finally announce one of the most ambitious projects we've worked on — which makes the front cover of Science Robotics today: ☀️ Learning a Thousand Tasks in a Day ⭐️ Everyday tasks — like those below — can now be learned from a single demonstration each...

VLA model developers like @LeRobotHF aren't concerned enough with robotic safety. They need some skin in the game. I would like to propose a new robotic competition format for ICRA 2026:

@KyleMorgenstein @observie Kyle, do you have a writeup / paper where this idea is presented in more detail? (using super low gains and letting the policy "muster" its own stiffness by generating large joint angles). I'm writing a paper on compliant policies and would be happy to cite something.

@observie for RL you want to use gains so soft it couldn’t support the robot’s weight under gravity. for model based control double the P gain. you should not have to adjust the PD params at runtime. run the PD controller as fast as you can (I’ve seen up to 40kHz) and interpolate commands

Impressions after spending a Saturday with RoboDog 1. Motors' behavior under load is a whole new ball game. 2. A position command that easily sends an unloaded motor to the desired position behaves completely differently when motors are loaded. 3. This is amplified when considering gravity: the same positioning strategy that struggles to stand up can result in abrupt collapse when lying down. 4. Positioning strategy is generally defined by PID parameters, step size from position to position, feedforward torque/velocity parameters embedded in position commands, and control loop frequency. 5. Should we consider changing PID params depending on desired dynamics? It might be complex during closed-loop control, though. 6. Example: for a gradual posture change, do we use mild proportional (P) gain and one position command, letting the PID control loop handle the entire change, or high proportional gain but split the movement into incremental position changes? 7. Need to find a systematic plan for selecting PID params - currently it felt too much like that DJ meme 8. Playing with the integral (I) term in the PID loop should be done cautiously: I broke a PA12 shaft on the first attempt. I currently tend to set the integral term to zero, essentially achieving a PD control loop, not PID. 9. IMU might be needed sooner rather than later: switching positions under load may be more successful if postures are defined in terms of physical orientation rather than motor position numbers. 10. Generally, spending time with this problem feels beneficial before switching to RL - getting immersed in this problem domain will probably make future debugging easier. @suren_at @KyleMorgenstein

@benjamin_bolte You probably have a base height tracking reward — you can gently raise the desired height to have during training (even simpler: just condition the policy on a small range of height goals)

@aditya_bhatt Trying out AMP. Had a hard time getting the robot to explore properly is the legs aren't set to bent by default

Messing around with a humanoid robot in your backyard will soon be a common occurrence. Policies seen here were trained in Mujoco, IsaacLab, and IsaacGym. There is no "EDU" version - everything here is unlocked, and always will be. Free as in Freedom

@cixliv Test the policy in an interactive mujoco simulation before you deploy. Hang it (click + drag), push it, perturb with the mouse cursor. I always do this to each checkpoint before deploying You *can* train policies with gentle responses to unforeseen situations - regularise!

the technical reason is we ran a full body policy while the feet weren't touching the ground. don't do that.

please make this go viral so I can pay for repairs our humanoid robot boy DeREK completely lost his mind @rek