Five iterations and six hours. Enough to fine-tune a humanoid motion generator with RL. 📌 For anyone working on humanoid locomotion or generative robot policies, this paper is worth saving. The PPO baseline used more than 160x the data and 30x the wall-clock time, and still ended up about 13 points lower in success rate. The work comes from Zachary Olkin, William Compton and Aaron Ames at Caltech's AMBER Lab. Their Unitree G1 plans 1.24 seconds of whole-body motion directly from depth images and replans every 0.24 seconds, while a separate tracker runs at 50 Hz. The generator produces 2,728 values for every plan, which makes ordinary action-space exploration extremely inefficient. Instead of adding Gaussian noise across the whole output, they perturb the conditioning, resample the initial flow-matching noise and start the robot from states the original policy would rarely visit. The nominal conditioning is kept as the label. A critic is trained on the generator's frozen prefix encoder and assigns advantages to the collected trajectories. Those advantages weight the flow-matching loss, giving better trajectories more influence on the next version of the generator. The reward is built around terrain consistency. Penetration and contact are checked first using signed distance fields, while velocity tracking and task success only count when the planned motion is consistent with the terrain. After five iterations, success on unseen geometries and skill combinations improved by as much as 25 percentage points. On one staircase, the share of runs that selected every required skill correctly went from 0% to about 80%. The ablations are useful here. Plain behavior cloning performs poorly, removing the conditioning perturbations costs around 4 to 6 points of success, and PPO on the full 2,728-dimensional output needs far more data while still performing worse. The camera setup also shows up clearly in the numbers. With both depth cameras, box success in simulation is 100%. With only the lower camera, it falls to 57.3%, because the upper camera gives the robot enough warning to slow down before a box or staircase and speed up again afterward. The full system runs without odometry or height maps. Depth images go into the motion generator, the tracker follows the generated motion, and the fine-tuning loop improves the generator from collected rollouts. Website: Paper: Video: —— Weekly robotics and AI insights. Subscribe free:

Intrinsic capacity isn’t simply fitness, independence or being disease-free. It spans cognition, locomotion, psychological and sensory capacity, and vitality, but there’s no universally accepted composite score yet. #longevity# #geroscience# #intrinsiccapacity#

Predator is a steel humanoid that got knocked to the floor repeatedly this week & stood back up unassisted every time. Fall recovery is harder than walking, & it's the part worth watching. Upright locomotion is a balance problem with a known starting state. A fall erases that. Predator lands in an arbitrary orientation, has to reconstruct its position from its own sensors, then sequence dozens of joints through a push-up while its center of mass moves the entire time. One error & it's back down. It did this under adversarial conditions, against Frankie LaPenna trying to prevent it. To be clear about what's being measured: REK's founder told NBC News these robots aren't autonomous. They are "100 percent" human pilots in VR headsets. So Predator is testing teleoperation latency, impact tolerance & balance recovery, not machine decision-making. Which makes it a more honest benchmark than most. Lab demos happen on flat floors under controlled conditions. You can't stage a knockdown from someone actively fighting you. Getting up is the gate on humanoids working anywhere near people. Predator cleared it in public, repeatedly, on video. 🦾

AGIBOT welcomed a Thai delegation led by Mr. Songsak Thongsri, Deputy Prime Minister of Thailand and Chairman of the Board of Executives of the National Sports Development Fund (NSDF), to our Shanghai headquarters. The delegation watched humanoid robot demonstrations showcasing locomotion, interaction and manipulation capabilities, and learned about real-world applications across industrial, commercial, cultural, sports and educational settings.

"Hey Alexa... uh, I mean Atlas, hand me the drill." Boston Dynamics has a new CEO: Rohit Prasad, the man who built Alexa. And it's a notable break from BD's past. BD's first two CEOs were hardcore roboticists who helped pioneer the field. Founder Marc Raibert built the legendary MIT Leg Lab and basically invented modern legged locomotion. Robert Playter spent decades deep in the dynamics and control behind Atlas. Robotics people, through and through. Prasad is an AI leader, not a roboticist. At Amazon he was SVP and Head Scientist for Alexa and AGI, and led the Nova foundation models. Before that, 14 years of ML research at Raytheon BBN. He's only BD's 3rd CEO ever. The mandate: push their "Physical AI" bet, fusing AI with the robots for industrial deployment, backed by owner Hyundai. For 30+ years BD led on mechanics and classical control. Now it's betting the next era is won on AI. Big shift.

A robotic hand that walks away on its own fingers! 🪬 We finally have the Thing from Addams family in robotics, and it was developed at ETH Zürich. No wheels, no legs bolted underneath, no separate drive mechanism. The same fingers that grasp things also carry the hand across a surface and hold up its weight. A quadruped has four identical legs spaced evenly around its body, and most legged-locomotion reward functions quietly assume exactly that. A hand has fingers of different lengths, arranged unevenly around a palm, with a thumb that matches none of them. The team wrote a reward formulation for that unequal geometry, and in simulation it moves faster than tuned rewards borrowed from quadruped work. They kept the hand's existing finger design and its position controller, so this isn't a bespoke machine built for the demo. And power and computation sit onboard, so it runs untethered. → Crawling, steering, and recovery after being knocked over → While holding itself up, it executes a sequence of keyboard commands with no vision at all → It pushes an object to target positions using overhead visual feedback → The simulator was calibrated against measurements taken from the physical hardware A hand that walks to a task and then performs the task is a mobile manipulator with no locomotion hardware in it. Big shoutout to authors @amrhkzp! 😮‍💨 🔗 Here's the paper: ~~  ♻️ Join the weekly robotics newsletter, and never miss any news →

TRON Camp 2026 Documentary EP.04 | When Code Walks Out of Simulation By Day 1, the students had already brought a humanoid robot to life. Across five days, they moved from simulation to real hardware, validating, deploying, and tuning their algorithms on real terrain. What once felt like a long journey from code to reality became a first step, taken by the robot — and by them. TRON Camp 2026 officially comes to a close. See you next year! #LimXDynamics# #TRONCamp2026# #EmbodiedAI# #Locomotion# #Sim2Real#

A dexterous hand that stands up and does pull-ups! 😮‍💨 This is Handroid that is a gripper that can transform into a little humanoid robot. It was developed across teams from The University of North Carolina at Chapel Hill and Stanford University. It looks like The Thing and I'm totally obsessed with it! 🫳🏼 It's a desktop-scale robot with 27 degrees of freedom that runs as a dexterous hand when manipulation is needed, and as a humanoid when locomotion is the prio. The secret sauce is in the morphology. The finger joints double as humanoid limbs, and the same compact actuator modules get reassigned across fingers, arms, legs and base actuation. P.S. Imagine your grippers being able to come down and start pushing things :D 🔗 Here's the project page: ~~ ♻️ Join the weekly robotics newsletter, and never miss any news →