Physical AI

SkeleWAM: squeezing a manipulation scene into a sparse 3D skeleton

World action models usually represent future state as video or visual latents, which encode geometry only implicitly and carry appearance irrelevant to control. SkeleWAM keeps just three kinds of 3D landmarks - robot joints, object centres and interaction points - and uses that sparse skeleton for both action generation and future-skeleton prediction, training with the latter as auxiliary supervision and dropping it at inference. On LIBERO-Plus it reaches 85.9% zero-shot success with 57.1M parameters, beating the 2B-parameter Cosmos-Policy by 3.7 points.

Juyi Sheng, Hua Wang, Mengyuan Liu

Boston Dynamics unveils new Atlas hands: 13-DoF directly actuated, built for high-fidelity simulation and sim-to-real RL

Boston Dynamics redesigned the hands of its electric Atlas: 13 degrees of freedom, directly actuated, designed from the ground up for high-fidelity simulation enabling sim-to-real RL, plus dexterous physical work — rugged, mass-manufacturable and repairable at low cost, with a form factor close enough to a human hand to train from human demonstrations. The previous three-finger hand performed superhuman manipulation but was never designed for tens or hundreds of thousands of units. The design philosophy: hands are a ruthless tradeoff, and many current designs give up reliability and manufacturability. Atlas (1.9m/90kg/56 DoF/50kg instantaneous lift) is heading into real factory work for the modern AI era.

Figure sends its retired F.02 humanoids leaping autonomously into a Finnish foundry furnace: a Terminator 2 homage whose real motive is IP-safe disposal

Figure AI gave its retired second-generation humanoid F.02 a Terminator-style send-off: with the newer F.03 fleet arrived, maintaining old units no longer made sense, and dismantling each machine would eat engineering time needed for the next-gen F.04 while risking exposure of proprietary hardware and IP. After multiple steel mills declined, only a foundry in Imatra, Finland took the job. The team rigged a large air cushion at its San Jose HQ, captured stunt-performer motion data, and trained a new AI motion model in simulation so the robots could **autonomously leap from a second floor and land precisely inside a 75-ton electric arc furnace they had never surveyed** — then be melted into steel. The company cut a 3+ minute video homaging the T-800's thumbs-up furnace scene in Terminator 2 (complete with the thumbs-up), but blogger xiaohu finds it "a bit tragic and cruel — I worry about the robots of the future seeing this." Context: before retirement the F.02 fleet spent 11 months at BMW Spartanburg — 30,000+ X3s assisted, 90,000+ sheet-metal parts loaded (84 s cycles / 37 s loads / 99%+ accuracy), 1,250+ runtime hours, ~200 miles walked, zero injuries — with those battle scars feeding directly into Figure 03's redesign (no more wrist dynamic cabling or distribution boards). Only a few F.02 units remain at HQ for display.

Mind-1: Human-Speed Physical AI for Real Robot Work

MindOn's Mind-1 combines human-speed data, low-latency inference, and whole-body control to run packaging, sorting, and folding across three robot embodiments.

Temporal cascading of planning and control in one quadrotor MPC: high-fidelity control nearby, point-mass planning far ahead, up to 75% lower closed-loop cost

A University of Zurich team (Rudolf Reiter, Chao Qin, Leonard Bauersfeld et al.; IEEE T-RO paper, arXiv:2512.12427) introduces temporal cascading for quadrotor MPC: instead of separating planning and control (a simplified model plans a long-horizon trajectory for a high-fidelity controller to track), both live in one MPC optimization — a high-fidelity quadrotor model controls the near future while a computationally cheap point-mass model plans the long horizon. The two horizons are not merely stacked but coupled through transition and feasibility constraints, so what the simplified model plans remains executable by the real quadrotor; parallel low-fidelity planners and progressive obstacle smoothing escape local minima in cluttered environments without slowing the real-time loop. Result: long-horizon planning with iteration times below 5 ms, and in simulation and real-world flights up to 75% lower closed-loop cost than standard MPC and conventional hierarchical planner-controller stacks under comparable compute. Narrated video: youtu.be/CefIqpld9N8.

OTRetarget: joint robot-and-object motion retargeting via optimal transport preserves contacts across morphologies

OTRetarget (arXiv:2609.36602) by the Simple Robotics team (Guillaume Besset, Erwann Carn et al.) jointly retargets robot and multi-object motion from human demonstrations: surface interactions are described by signed distances, closest witness points, and relative directions; one entropic optimal-transport plan per human part and robot link (solved once in a T-pose) carries these targets across morphologies so fingertips and heels are not left behind. Interaction targets feed a constrained inverse-kinematics formulation that jointly optimizes robot and every object pose per frame, balancing contact preservation with motion style — robot-object and object-object interactions without rescaling the scene. On OMOMO it reaches 87% robot-object interaction Jaccard and 8.7 mm depth error vs OmniRetarget's 28% and 29.3 mm; one demonstration replays onto unseen object shapes (ball to beam), and a policy blind to the object completes pick-and-place on a physical Unitree G1.

AMB3R-SLAM: kilometer-scale real-time SLAM on a single consumer GPU, no bundle adjustment needed

AMB3R-SLAM (arXiv:2609.19518) by Hengyi Wang and Lourdes Agapito at UCL: a training-free, model-agnostic real-time monocular SLAM that reconstructs kilometer-scale trajectories over 10k frames on a single consumer GPU. A hierarchical backend (span-2, long-context, and loop-closure pose-graph edges) progressively enforces local, mid-level, and global consistency while avoiding bundle adjustment and its static-world assumption, so dynamic scenes work out of the box; the same framework extends to stereo, RGB-D, and LiDAR inputs. Across 9 datasets it cuts ATE of prior state of the art by over 70% on VBR and Oxford Spires, and reaches sub-meter ATE on KITTI and VBR with LiDAR input. Shared by Ryohei Sasaki, localization & mapping engineer at MAP IV (TIER IV group).

Goose home robot debut: not humanoid - a low body plus long articulated neck takes another shape for the home

San Francisco startup The Robot Works unveils Goose, prototype 1 of a household robot that deliberately avoids the humanoid form: a white, rounded, low base close to the floor, a long articulated neck reaching outlets, door handles, counters and laundry piles, camera eyes set in a bonnet so you can see where it looks, an orange beak doubling as a gripper, and a blue polka-dot wrap making it feel like an appliance rather than a machine. After interviewing many families, founders Richard Wei and Humphrey Hu (ex-Anki, expressive robot faces) chose this shape: a wide low base is safer than legs, and a neck suits reaching better than standing tall; motors yield when blocked by a foot, chair or child, and the robot analyzes blind spots before moving. Demos include placing a foot on an ottoman, gathering clothes off the floor, reaching the kitchen counter, plugging into a wall outlet, and opening a door for arriving guests. The business starts with services: full-room tidy, checking in on grandpa, story-time companion, home monitoring while away, stove checks, and bringing parcels in - positioned as an assistant that works alongside you, explicitly not doing hands-on care like bathing, dressing or lifting.

ETH RSL: high-speed rough-terrain locomotion for ANYmal via automatic curriculum RL

ETH Zurich Robotic Systems Lab (Marco Hutter group, RAL 2026) presents LP-ACRL: a Learning Progress-based Automatic Curriculum Reinforcement Learning framework. It estimates the agent learning progress online and adaptively adjusts the task-sampling distribution, requiring no prior knowledge of the difficulty structure over the task space - solving the long-standing problem that difficulty ordering is undefinable in complex, wide-ranging task spaces. Policies trained this way let the ANYmal D quadruped sustain 2.5 m/s linear velocity and 3.0 rad/s angular velocity across stairs, slopes, gravel and low-friction flat surfaces, where previous methods were limited to high speeds only on flat terrain or low speeds on complex terrain. The paper highlights scalability and real-world applicability, offering a robust baseline for curriculum generation in complex robotic learning task spaces.

Q2: an RL-first quadruped redesign with four identical legs and direct-drive knees

The author presents Q2, a redesign of their Q1 quadruped, and describes it as a major simplification and an AI-first design: the layout was first checked to reach fast RL convergence in simulation, and only then moved into CAD. The eight listed points are four identical legs with no left or right parts, direct-drive knees with no linkage or belts, no reduction gear after the motor output, knees that rotate 360 degrees so there is no predefined knee posture, a low standing load that needs almost no knee torque, shorter links for lower leg inertia and joint torques, the battery placed at the geometric center for a better center of gravity, and a single spec file that drives both the simulation model and the CAD. The author calls this parametric CAD taken to the extreme, where the CAD software is only used to review the design rather than to build it, and says weight, cost, manufacturing time and assembly effort would all drop significantly while the battery design, PCB and compute stay the same.

Systematically Exploring the Capabilities of GPT-6 Astra as Embodied Policies

Evaluates GPT-6 Astra as an embodied policy across six domains: hybrid control reaches 48% on RoboDojo and 92% on RxR navigation, while direct in-hand control and dense locomotion references stay unreliable; token use and inference latency constrain practical control.

Galbot Team, Chen, Xuchuan, Cheng, Xiaoqian

Delta Intelligence unveils Delta-0: a humanoid foundation model for whole-body loco-manipulation across 69 DoFs

Delta Intelligence has released Delta-0, a humanoid foundation model for whole-body loco-manipulation built on a co-designed brain and controller. The brain is a latent world-action model: a mixture-of-transformers with three branches (vision-language reading multi-view observations and instructions, a vision branch predicting future frames in DINO semantic feature space, and an action branch modelling whole-body motion from proprioception), trained jointly in four modes - forward dynamics, inverse dynamics, visual planning and policy-only. The controller converts motion commands into joint targets across 69 degrees of freedom through a delta-action interface shared by policy output, teleoperation and human corrections; in zero-shot evaluation on roughly 3 h 31 min of everyday-task motion it beats HEFT, MimicLite v1.1, SONIC v1.1 and ScaleBFM XL on both global root error and MPJPE. Pretraining uses more than 10,000 hours of paired egocentric video and whole-body human motion, plus a 154-dimensional shared action representation padded to 180 so single-arm, dual-arm, ego, UMI and humanoid data all live in one action space; high-precision tracking coverage rises from 29.2% to 72.6% as motion data goes from 10% to 100%. Evaluation runs a real-to-sim-to-real loop, rebuilding deployment scenes into simulation assets with 3D reconstruction and LLM agents, running repeatable autonomous rollouts, then returning selected checkpoints to hardware. Real-world RL post-training uses a value model that estimates cost-to-go, giving positive or negative conditioning signals to autonomous action chunks and to human-in-the-loop corrections - on the dishwasher task in a mirrored kitchen layout the success rate climbs from 20% (4/20) after SFT to 65% (13/20). The demo video covers six long-horizon household tasks: loading a record on a turntable, making a bed, picking a toy off the floor, opening a dishwasher, pressing a step trash can and sitting down on a sofa.