NEW SKILL: This team take humanoid parkour to the next level! This project involves @zhenkirito123, @Yuanhang__Zhang, @pabbeel, @carlo_sferrazza, @GuanyaShi, and others. Called Perceptive Humanoid Parkour (PHP), it lets a Unitree G1 humanoid (29 DoF, 1.3 m) run long-horizon, vision-based parkour and decide on its own whether to step over, climb onto, vault, or roll off obstacles, using only onboard depth sensing (a 30 Hz camera) plus a discrete 2D velocity command from the operator. It builds the motions by motion matching, a nearest-neighbor search in a 27-dimensional feature space (imported from the video-game industry) that stitches retargeted atomic human clips into long-horizon kinematic trajectories through a shared Locomotion to Skill to Locomotion manifold. It then trains per-motion RL tracking experts (with a privileged height scan) and distills them into one depth-based multi-skill student via DAgger combined with RL. It replaces hand-scripted skill sequencing and operator-triggered skill selection. The onboard depth-only student holds 0.95 to 1.00 success across obstacle heights in simulation, while a naive end-to-end depth policy collapses from 0.95 to 0.07/0.08 at 58 and 76 cm. The upside comes from distilling privileged-height-scan experts and from motion-matching composition, not from training onboard from scratch. The operator sends only a coarse 2D velocity (W/A/D keys), and the robot itself picks which skill to run from the depth image. The training data contains only single-obstacle traversals, so the ability to chain skills across a multi-obstacle course is generalization the policy was never explicitly taught. Transitions are constructed in kinematics, not rediscovered by learning. Every skill enters and exits through a shared locomotion manifold, so no hand-captured skill-to-skill transition clip is needed. Motion matching densifies a sparse library into smooth long-horizon trajectories before any RL. The entire parkour library is only about 66 seconds of mocap, and each skill is just a few seconds. Ablating approach-distance variety ("Extreme Distances") drops the 76 and 94 cm climbs to 0.62 and 0.64, and "Half Density" drops the 76 cm climb to 0.32. Motion matching's job is to manufacture stride-phase and approach variety out of seconds of data. It clears a 1.25 m wall (96 percent of its 1.3 m height) in 3.63 s, and a cat-vault peaks at 3.41 m/s clearing a 0.4 m by 0.5 m obstacle in 0.8 s. A single 30 Hz depth stream drives both traversal and runtime re-planning, shown by displacing obstacles about 0.5 m mid-run.

NEW RESEARCH: You can play soccer with your Unitree G1 Humanoid! This project involves @PeterYChong, @YuhangLu2000, @sephy_li, and others. It is a collaboration between OpenDriveLab, the University of Hong Kong, the Chinese University of Hong Kong and Archon Robotics. Called RoboNaldo, it is a reinforcement-learning policy that makes a Unitree G1 humanoid (29 DoF, 35 kg) shoot a soccer ball, both stationary free-kicks and one-touch shots on a moving ball. It is trained by a three-stage motion-guided curriculum in Isaac Lab (4096 parallel environments): - Stage 1 imitates a single human-kick reference (retargeted from a human video via GVHMR and GMR) with no ball or task reward - Stage 2 adds the ball, a target and shooting rewards while randomizing the ball spawn - Stage 3 adds a high-level locomotion command and a kick trigger so it can walk up and volley Perception is onboard and egocentric: a head-mounted Livox MID-360 LiDAR and a chest RealSense D435 track a retro-reflective ball (Kalman-fused), and an AprilTag (or a fixed coordinate) marks the target, all at 50 Hz with no motion capture. What is interesting imho: if you remove Stage 1 (imitating one retargeted human-kick video), the alive rate collapses to 1.2 percent, it topples on nearly every attempt. The human prior's real job is whole-body stability through a violent, high-impulse motion, not style; the hard part of a robot kick is not aiming but not tipping over from your own swing The onboard, no motion capture perception still needs a retro-reflective ball and an AprilTag goal. Ball localization is a head Livox MID-360 LiDAR plus a chest RealSense IR camera tracking a retro-reflective ball, fused by a Kalman filter. The target is read from an AprilTag or hardcoded as a fixed coordinate. It shoots hard but sub-human and only roughly accurate. Peak ball velocity is 13.10 m/s, which it reports as 59 to 71 percent of a professional open-play shot (about two-thirds of a real Ronaldo, fitting the pun). Stationary accuracy is 0.73 m average error from 3 m, with 31.5 percent of shots inside 0.5 m and 80.6 percent inside 1.0 m. The entire motion prior is one human video. There is no kick dataset and no mocap suit: a single human-kick clip retargeted to the 29-DoF G1 via GVHMR and GMR, then curriculum RL across 4096 Isaac Lab environments deviates from it for aim, contact and timing. It is the same one-demonstration-plus-sim-compute recipe as LadderMan and the perceptive-parkour projects I talked about previously. Also fun fact: it is an autonomous-driving lab pivoting to humanoid soccer. RoboNaldo comes from OpenDriveLab (the HKU end-to-end self-driving / UniAD group) with CUHK and the startup Archon Robotics, all Hong Kong based.

NEW ROBOT SKILL: Your humanoid can now play table tennis with you! It involves @TaouHuang, @junli_r84995, @kai_zhangjojo, @sephy_li, @ilnehc, and others. Called SMASH, it lets a Unitree G1 humanoid play table tennis from its own onboard cameras, with a single whole-body RL policy (no split upper/lower body). A head-mounted ZED X stereo camera tracks the incoming ball at 60 Hz. A downward-tilted ZED X Mini localizes the robot at 120 Hz by reading seven AprilTags placed on the table, and an Adaptive Extended Kalman Filter with a quadratic-drag ball-flight model fuses the two async streams into a joint estimate of ball state and robot pose at over 50 Hz. The policy is trained on a strike library of 400 real motion-capture demonstrations expanded by a motion VAE to about 5,000 clips, and produces coordinated full-body shots including smashes and low crouching returns. It replaces prior table-tennis humanoids that split upper and lower-body control and depended on external sensing. "No external cameras" does not mean no instrumentation: self-pose still depends on seven AprilTags on the table. The robot carries the cameras, but the downward ZED X Mini localizes it at 120 Hz by reading those printed fiducials (AprilTag plus PnP-RANSAC). The infrastructure moved from active (external cameras or mocap) to passive (printed tags) -> it did not disappear, so "onboard only" is softer than it first sounds. The honest return rate is 59.7 percent, not the headline 93.7. Over a 50-minute, 642-launch run, 93.7 percent is the contact rate, but 34.0 percent is contact-only, so about 40 percent of contacts do not come back over the net, successful returns are 59.7 percent. The backhand is the clear weak side: 38.9 percent return versus 66.7 forehand, with 52.1 percent of backhand contacts failing to return. The impressive number is contact; the useful number is return. In simulation it ties the HITTER baseline on success (86.38 vs 86.63). Its measurable edges are motion quality (whole-body MPJPE 75.0 vs 100.1) and the real-world capability HITTER lacks: consecutive rallies outdoors on onboard sensing alone. The follow-up beats the benchmark by moving to onboard egocentric vision, not by scoring higher on the shared metric. Using the head ZED X for pose instead of the downward ZED X Mini drops returns from 13/20 to 8/20; removing the adaptive observation-noise scaling drops to 7/20; and ablating the YOLO detector with a yellow distractor present takes detection to 0/20. The self-pose channel, not just ball tracking, is what makes returns land. One caveat: no measured smash, ball, or paddle speed appears anywhere, so "explosive whole-body smashes" is qualitative and the agility claim is unquantified in velocity terms. Besides, the project's claim "100-plus consecutive rallies" is marketing: the paper counts 642 ball launches with randomized landings, not 642 completed rallies.

NEW ROBOT: Origami wall climber! Called SPARC, it is a collaboration between Shanghai Jiao Tong University's State Key Lab of Mechanical Systems and the University of Michigan's Hybrid Dynamic Robotics Lab. The robot here is a 210 g, 160 mm soft crawling/climbing robot driven by three parallel 3D-printed vacuum-powered Kresling origami actuators, with four silicone suction cups for adhesion. A Kresling origami is a fold pattern that turns a thin cylinder into a spring-like structure of twisted triangular facets. Compressing it makes it rotate and snap between stable states, so engineers use it to build lightweight actuators and robot bodies that fold, twist, and lock without motors at every joint. Each actuator contracts under vacuum, and because Kresling origami mechanically couples that contraction one-to-one with a twist angle, a magnetic angle encoder (AS5600) at each actuator reads the twist and an inverse-kinematics model recovers the shape, giving the robot proprioception. A dual closed-loop controller (onboard encoder/PID inner loop plus an external motion-capture outer loop) steers it along horizontal and vertical paths, and two modules in series can bend about 100 degrees to cross from floor to wall. The thread pairs it with a quoted post on Festo's BionicMotionRobot, a positive-pressure pneumatic-bellows soft arm. Soft robots normally cannot sense their own configuration because stretchable strain sensors are noisy and drift. SPARC sidesteps that with geometry. Kresling origami couples axial contraction one-to-one with a twist angle, so a cheap magnetic angle encoder (AS5600) reading the twist plus an inverse-kinematics model recovers the length. Proprioception becomes a geometry problem solved with a few-dollar off-the-shelf encoder, not a stretchable-materials problem, and that is what lets a soft robot hit 0.5% trajectory error, a precision usually reserved for rigid machines. The controller is a dual closed loop: the inner loop is the onboard encoder/PID (the genuine proprioception), but the outer loop is an OptiTrack Prime41 motion-capture system feeding global position. So the 0.5% horizontal / 3% vertical figures are achieved with off-board global tracking, not self-contained navigation, strip the mocap and the robot still knows its shape but not where it is. Same external-infrastructure caveat this file keeps flagging (SMASH's table AprilTags, LadderMan's bench-side GPU): the impressive number quietly depends on lab instrumentation. Vacuum actuation needs an off-board air supply through proportional electrovacuum regulators (SMC ITV209), an OptiTrack rig provides localization, and there is no onboard power (an onboard Jetson Nano and Arduino Mega only run the valves and encoders). It is a demonstrator of the sensing-plus-actuation idea, not a deployable crawler. Each Kresling actuator contracts 60% at -80 kPa with about 3 kg of actuation force, and three of them on a 210 g body is why it climbs a vertical wall carrying 500 g, roughly 2.4 times its own weight. Vacuum actuators collapse rather than burst (inherently safe) and pack high force, whereas Festo's BionicMotionRobot inflates twelve positive-pressure bellows to bend an octopus-like arm. Same "soft robotics" banner, opposite pressure sign and opposite job (mobility versus manipulation). The ground-to-wall transition currently takes two robots, not one. A single module bends about 50 degrees; two connected in series reach about 100 degrees and double the step length, which is what lets the pair round the 90-degree floor-to-wall corner. Autonomous single-module ground-to-wall transition is explicitly listed as unsolved.

.@Moca_Network, our flagship project, has signed an MoU with @InveoKripto and Ichain Investment Holding to advance regulated Web3 infrastructure across Türkiye. At the core is AIR, built by Moca Network: a privacy-preserving SDK on three pillars — Identity, Money, and Loyalty — embedded directly into the apps people already use. One identity. Infinite ecosystems. Full announcement:

Fwiw $BUN mcap on dexscreener / gecko are wrong. Only 20% float rn since ~80% supply is permanently locked So real mcap is 0.2 x price, which currently equals ~$16M instead of $22M on DS has the right numbers

double digit IQ 20k mcap shitcoin traders did 5 minutes research and have something to say about my $BEM position “oH bUt ItS mInEaBle” yes motherfucker, that’s literally why i’m bullish on it.

Market Overview: 🔹 Crypto mcap down 4.26% to $2.47T. BTC -4.52%, ETH -13.57%. 🔹 Long liquidations hit $654M on Thursday, nearly double the usual. 🔹 Funding rates rising across majors despite the drawdown. 🔹 Strategy buys zero BTC this week, repurchases $1.5B in 2029 convertible notes instead. 🔹 MasterCard secures NY BitLicense for stablecoin and digital payment infrastructure. 🔹 Grayscale delays its IPO to at least Q4 2026, citing market conditions. 🔹 S&P +2.06%, Nasdaq +3.68% as equities soared while crypto failed to find a bid. 2/6