LightOrigins

@irvinxyz Lovely :)

Good morning to everyone working on robotics / Physical AI☀️🤖 Don’t forget to touch some grass.

@jiayu33505 Thank you!

@LightOrigins_ so cool

One Policy. No Odom. Depth Only.

@HailinXion27236 Thank you!

@LightOrigins_ It looks absolutely amazing!

@Prometheacc on a mission :)

@LightOrigins_ Where is he going ?

@ErenChenAI Thank you！

@LightOrigins_ Impressive work

@sebastianboehle Thank you! Sim training and then real-world deployment. We will post our paper very soon!

@LightOrigins_ What is your sim and training stack? Looks awesome!

@carlosdponx Only forward 😀

@LightOrigins_ very cool! forward depth only? or does it have a backward facing one too?

@WerthFinance Thank you!

@LightOrigins_ A+ for those sound effects. Nicely done.

@TheSeekerOf42 @ErenChenAI The thought of this research is that in future large-scale deployments, unexpected situations are inevitable. We want the robot to still be able to adapt when those situations occur.

@ErenChenAI Pretty impressive to cram all that movement modes into a single unit! But what’s the use case? Unless it’s some combat/hazardous applications for a robot, it’s more a PR trick than practical feature.

A new Chinese robotics startup, Light Origin, just showcased an impressive single-policy locomotion system. Even with one or both legs restrained, the robot remains mobile by hopping or crawling.

@Tanovski @ErenChenAI I will post it on my page hopefully this week :)

@ErenChenAI Is there a paper?

@clankrmedia @RogerXJiang It doesn’t need to be selected manually. It can automatically recognize the situation from history and adapt accordingly.

@LightOrigins_ @RogerXJiang Looks really cool. One question though: can the robot recognize which attachment is connected, or does it have to be selected manually?

We tied our robot up. One policy walks, jumps, and crawls. Restrain it and it finds its own way back to stable walking from a state we never scripted. Behind it is our founder @RogerXJiang , an ex OpenAI scientist who worked on RLHF, the method behind ChatGPT. We're applying that same idea to robots: train one general policy instead of scripting each move. Follow us for more.

@chris_j_paxton Yes!! In large scale real world deployment, unexpected situations are inevitable. We want our policy to handle them, not just fuction in a controlled lab environment.

Fundamentally this is why wheeled robots will lose in many environments (over a 5-10 year timeline). In the end you want a robot that can always stay operational, and there are so many ways a legged robot can adjust and stay stable.

@clankrmedia This is too cute!

Researchers built a soft floating robot for indoor interaction. It uses helium and flapping fins instead of propellers. The result is quiet, lightweight, and safe to touch. It can follow people, give reminders, and act as a study buddy. Published at ACM DIS 2026.

@Em_Nomadic 🧠 clever? yes. body clever? also yes.

Morphological Computation Here’s something that quietly unsettles our whole idea of intelligence. Plants solve problems, and they have no brain at all. A root growing through soil isn’t drifting. It senses moisture, gravity, nutrients, toxins, and the presence of other roots, and navigates around obstacles toward what it needs. Researchers have identified at least 22 distinct signals a plant can sense and respond to. And it can be genuinely sophisticated. A 2017 study found that Arabidopsis seeds decide when to break dormancy using what the authors called a spatially embedded decision-making center, a distributed computation spread across cells, weighing temperature history to pick the right moment to sprout. No neurons. The computation lives in the tissue. Charles Darwin saw this coming. In 1880 he proposed the root-brain hypothesis, suggesting the tip of a root behaves like the brain of a simple animal. The idea sat forgotten for over a century and is now returning as molecular biology reveals plants as sensing, signaling, problem-solving organisms. The plant neurobiologist Stefano Mancuso adds a striking twist: plants have no brain not because they lack intelligence, but because a brain is a vulnerability. Centralized intelligence is a single point of failure. A plant can lose 90 percent of its body and survive, because its intelligence is everywhere. We can’t lose a fraction of our brainstem. Evolution gave plants a different answer to the same problem. Don’t put the mind in one place. This is where it connects to a genuinely radical idea in robotics called morphological computation. The claim is that a well-designed body can do part of the “computing” a controller would otherwise have to do. The body isn’t just hardware the brain drives around. It’s a computational resource. The purest example is a machine called the passive dynamic walker. The Cornell version walks down a gentle slope with no motor, no sensors, no code at all. Powered only by gravity. The stable walking gait, the balancing, the leg timing, the coordination, none of it is computed by a controller. It’s embedded in the physical structure itself, in the tuned lengths of the leg segments, the mass distribution, the curve of the feet. The “computation” required to walk was solved by the shape. It shows up everywhere once you look. A gecko’s foot clings to a vertical wall because of its morphology. No matter how big a brain you gave it, without those feet it would simply fall. Researchers built a soft silicone octopus-inspired arm that could perform complex computations through the dynamics of the material alone. Soft grippers passively conform to a strawberry or a wrench without any precise control, because the compliance of the body does the adapting. In each case, work that a traditional robot would burn computation on gets offloaded to physics. There’s real scientific debate about how far to push the word “computation” here. Some argue it’s more precise to say the body facilitates control rather than literally computes. But the deeper point survives the debate. Intelligent behavior does not have to originate in a central controller. It can emerge from the coupling of body, brain, and environment, with the body carrying real load. Which reframes the whole question of what intelligence even is. We inherited a picture where the brain is the thinker and the body is a puppet it operates. Plants break that picture. Passive walkers break it. Soft robots break it. Maybe intelligence was never a thing that minds have. Maybe it’s a thing that bodies do, in constant conversation with their surroundings, and a brain is just one especially concentrated, especially fragile way of doing it. The plant has been solving its way around rocks for 400 million years without a single neuron. It might have something to teach us about how thinking actually works.

@ritu_twts Impressive：）

"What if AI replaces you?" Me: I'll start farming. 😎 Meanwhile AI:

@spaceandtech_ Spider woman

Jizai Arms is a wearable robotic backpack from the University of Tokyo that lets users control up to six extra robotic arms through body movements. The system could improve work efficiency and accessibility across many fields. It was first introduced in 2022.

@TasbolatT Tas on the team 👏

Humanoid robots excel at long-tail tasks due to their redundant degrees of freedom, allowing them to adapt and continue performing tasks even when some joints or motors are impaired.

@techniahqrobot @LightOrigins @liangpan_t Thanks for the spotlight!!🙂We are showcasing that a single policy, the same one that walks and jumps finds its own way back to stable walking from a state we never scripted.

Most humanoid robots are trained to walk under clean conditions. @LightOrigins is testing what happens after everything goes wrong. The challenge was simple: tie the robot up and see whether it can escape. The robot crawls reorganizes its body recovers its posture and returns to locomotion. one movement policy controls walking jumping and crawling instead of relying on separate controllers for each behavior. The impressive part is the transition from restraint and crawling back to stable movement. Could one general movement policy make humanoid robots far more resilient in the real world? #HumanoidRobots #Robotics #PhysicalAI @LightOrigins_ @liangpan_t