Ajay Subramanian

Final run: success. After those fixes, the policy learns a much cleaner and more stable kick. The robot is no longer just matching frames; it is executing the motion in a physically consistent way. Code here: github.com/ajaysub110/rl_…

The fixes were mostly about getting the control formulation right: -initialize from the correct root body -include the reference motion in observations -predict residual actions instead of absolute joint targets -keep reward scales numerically sane -add balance/smoothness terms

As a fun pre-weekend project, I built a small humanoid motion imitation pipeline in Isaac Lab: take a human kick from AMASS (amass.is.tue.mpg.de), retarget it to Unitree H1 with ASAP (agile.human2humanoid.com), then train a PPO policy to reproduce it in simulation.

Excited to share two updates: I’ve defended my PhD at MIT, and I’ve moved to SF to join @Latent_Labs as a Member of Technical Staff. If you are in the Bay Area, I'd love to reconnect or meet up.

Why does manipulation lag so far behind locomotion? New post on one piece we don't talk about enough: The gearbox. The Gap You've probably seen those dancing humanoid robots from Chinese New Year. Locomotion isn't entirely solved; but clearly it's on a trajectory. But we haven't seen anything close for manipulation. 𝗪𝗵𝘆? When sim-to-real transfer fails, the instinct is to blame the algorithm. Train bigger networks. Crank up domain randomization. Those approaches have made real progress; we don't deny that. But we started wondering: are we treating the symptom or the disease? The Hardware Bottleneck: Fingers are too small for powerful motors. So most hands use massive gearboxes (200:1, 288:1) to get enough torque. But those gearboxes break everything manipulation needs:   • Stiction and backlash are complex to simulate. Policies trained on smooth physics hallucinate when they hit that reality.   • Reflected inertia scales as N². At large gear ratio, the finger hits with sledgehammer momentum.   • Friction blocks force information. The hand becomes blind. And they're the first thing to break. What we are trying to build at Origami, we cut the gear ratio from 288:1 to 15:1 using axial flux motors and thermal optimization. The transmission becomes more transparent: backdrivable, low friction, forces propagate to motor current. Early signs are encouraging. Still running quantitative benchmarks. Why Interactive? I love how Science Center uses interactive devices to explain complex ideas. I want to borrow this concept and help people understand the hard problems in robotics better visually. The post has demos where you can toggle friction, slide gear ratios, watch the sim-to-real gap widen in real-time. What's inside:   • Interactive demos (friction curves, N² scaling, contact patterns)   • Comparison table: 14 robot hands by sim-to-real gap and force transparency   • The math behind why low-ratio matters Read it here: origami-robotics.com/blog/dexterity… We're not claiming we've solved dexterity. The deadlock has many pieces. But we think this one's foundational. Curious what you think.

Thrilled to announce the final preprint of my PhD! We introduce PackFlow, a flow matching method for generative molecular crystal structure prediction, and post-trained via reinforcement learning on MLIP energies and forces. Paper: arxiv.org/abs/2602.20140 @RGBLabMIT

We’re making great progress with our Gemini Robotics work in bringing AI to the physical world - a critical aspect of AGI. As part of our next steps, super excited to announce our partnership with @BostonDynamics, combining our SOTA robotics models with their world-class hardware

Google DeepMind 🤝 @BostonDynamics Our new research partnership will bring together our advancements in Gemini Robotics’s foundational capabilities to their new Atlas® humanoids. 🦾 Find out more → goo.gle/49paguA

@simar_kareer @physical_int Congrats on the great work! The human data you use during finetuning is task-specific. Have you tried using cross-task human data, say as pretraining phase 2? Diverse robot data pretraining -> Diverse human data pretraining -> Task-specific finetuning. Would that help further?

Some properties of LLMs only emerge with scale, one of which is the ability to effectively generalize from diverse data. During my internship @physical_int, we uncovered an emergent property of VLAs: as we scale up pre-training, VLAs can naturally learn from human video data!

Wow! Scaling-up pre-training data in VLAs enables learning from egocentric human data. LLM-like patterns in robotics!

🚨 Appearing as a #NeurIPS2025 D&B spotlight(~3%) Could VLMs guess your next prompt for a wearable AI agent? We present WAGIBench, the 1st large-scale Goal Inference Benchmark for Wearable Agents w/ audiovisual, digital & longitudinal context! Paper: arxiv.org/abs/2510.22443 1/

@ky__zo Because 1) the world is designed for a human form factor, and 2) with current methods that use human demonstration data, robots that don't resemble humans are very hard to train

can someone explain why are we trying to make robots looking like humans? shouldn’t it have a crazy versatile form with 8 hands or legs, being agile and nimble to and light? it feels like closing the dishwasher would be easier if it was done by leg? (I almost always do it with my leg) forcing a robot to squat to close a dishwasher is like deploying a trpc turborepo for a static hello world page

Excited to present EclipseTouch at UIST today! 3:30 PM in Capri in the Sensing Light, Supporting Touch session. EclipseTouch is an accurate, headset-integrated system for sensing ad hoc touch input on a variety of everyday surfaces. Our system combines a computer-triggered camera with one or more infrared emitters to create structured shadows, which we can use to estimate touch contact and hover distance. Our results show this approach can be accurate, and robust across diverse materials, lighting conditions, and interaction orientations.

Everyone posting about the Dwarkesh interview (including Dwarkesh himself!) is missing this subtle point. When LLMs imitate, they imitate the ACTION (ie the token prediction to produce the sequence). When humans imitate, they imitate the OUTPUT but must discover the action

@olivia_y_lee Cool paper! Have you tried comparing your work in simulated environments against Eureka-style or RL-VLM-F-style reward functions?

Enabling robots to improve autonomously via RL will be powerful, and dense shaping rewards can greatly facilitate RL. Our #IROS2025 paper presents a method leveraging VLMs to derive dense rewards for efficient autonomous RL. ⚡🦾 #Robotics #ReinforcementLearning 🧵1/5

@alopeze99 Cool paper! Have you experimented with using a LLM/VLM to decompose tasks into stages rather than specifying it manually?

🤖Introducing DEMO3: our new model-based RL framework for multi-stage robotic manipulation from visual inputs and sparse rewards. 🧵🔽 📜 Paper: [arxiv.org/abs/2503.01837] 🌍 Project Page: [adrialopezescoriza.github.io/demo3/] 💻 Code: [github.com/adrialopezesco…]