Physical AI

WaterfallPagination

VIDEOSep 27, 2026

ROBOTIS Mission Canvas: Composing Reusable Robot Skills into Repeatable Mobile Missions

ROBOTIS released Mission Canvas, the mobile workspace of Autonomy Studio: define waypoints on a map, specify their order, and bind a Task to each location so navigation and robot actions execute as one Mission. The relationship to the earlier Action Canvas video is direct — Action Canvas defines what the robot does within a work area (VLA whole-body control, stationary manipulation); Mission Canvas extends that to where the robot should go and what to do there. Tasks can be deterministic robot actions or AI/VLA-based behaviors, so reusable robot skills compose into repeatable workflows across multiple real-world locations. Together the two videos show the progression from building individual tasks to deploying them as a mobile, repeatable Mission. Fully open source (Apache 2.0, cyclo_intelligence, 33 stars) with code, tutorial videos and technical docs for reproduction and extension. Docs note current support is limited to the ffw_sg2_rev1 robot type; Autonomy Studio combines autonomous navigation, Behavior Trees and VLA models in one visual tool — validate in simulation, then deploy to a physical AI Worker.

VIDEOSep 27, 2026

Sharpa Robotics Demo Draws Crowds: The Robot Plays Better Than Me

A repost praising a @SharpaRobotics demo: the challenge is open for anyone to try, and the author admits the robot is better than him (#robotics #physicalai). For context, Sharpa is a Singapore embodied-AI company behind North, a 67-DoF full-body humanoid, and Sharpa Wave, a 22-DoF five-finger dexterous hand with 1,000+ tactile pixels per fingertip, sub-mm resolution and 0.02 N sensitivity. At CES 2026, North ran script-free autonomous demos four days straight, eight hours a day: ping-pong rallies with a 0.02 s reaction time, ~2 mm-precise photo capture, card dealing via multimodal reasoning, and a 30+ step paper-windmill assembly.

VIDEOSep 27, 2026

AHa-3D: Real2Sim with GPT-6 Astra via Agentic Tool Use

Congrong Xu et al. found GPT-6 Astra alone misses objects and gets sizes and positions wrong when rebuilding a room video into a Blender scene. They built a harness handing Astra 3D tools — Pi3X for camera poses and metric-scale geometry, SAM3 for segmenting people/floors/walls, TSDF fusion, a measurement tool, the RoomKit asset library, X-Ray outline checks, and Bullet physics validation — organized as an Observe-Build-Verify-Interact-Deliver workflow. At similar cost and time, reconstructions are markedly better than using Astra directly, yielding interactive browser scenes with doors, furniture, materials, and human motion.

VIDEOSep 27, 2026

LimX Updates Cosa 0.5 to Give Humanoids More Intelligence

LimX Dynamics ships Cosa 0.5, bringing more intelligence to its humanoid robots with improved perception and task execution.

VIDEOSep 26, 2026

ROBROS IGRIS-C Demonstrates Human-Robot Collaboration

ROBROS introduces IGRIS-C, a humanoid robot PoC that moves two boxes to a conveyor belt one at a time and repeats the cycle as a worker returns them, highlighting safe coexistence in human-designed workspaces.

VIDEOSep 26, 2026

Proxy Policy Steering: Test-Time Robot Policy Adaptation

Proxy Policy Steering steers a frozen diffusion policy at inference time with a calibrated velocity residual from two lightweight proxies, raising average success by 53% across 12 manipulation tasks without modifying the base model.

PAPERSep 25, 2026

Generate, Track, Improve: Perceptive Multi-Skill Humanoid Locomotion with RL-Fine-Tuned Motion Generators

General purpose humanoids require locomotion controllers that are multi-skill, perceptive, dynamic, and robust enough to go anywhere humans can. In this work, we present a two layer locomotion architecture: (1) a perceptive flow matching motion generator plans whole body trajectories from raw depth images while a (2) perceptive tracking policy trained with control-guided RL follows these motions. Both policies are trained on a library of terrain consistent motion clips created with dynamically optimized human data which yields both accurate velocity tracking and terrain consistent references. Our central contribution is a simple yet effective off-policy RL fine tuning loop that improves the motion generator. A structured search method is used with the generator to gather data for advantage weighted regression. This off-policy loop is much more sample efficient than on-policy residual fine tuning and improves terrain consistency on unseen geometries and skill compositions. We find that successful terrain traversals increased by up to 25 percentage points and skill selection improved by up to 80 percentage points. By using raw depth images to perceive the environment no odometry or height maps are needed, and outdoor deployment is easy. With two cameras, the policy can see terrain coming from further away and adjust its velocity regardless of the commanded speed so it can traverse the terrain. A single policy pair enables a Unitree G1 humanoid to walk, run, stand, jump on and off of boxes, and traverse stairs in outdoor environments. Project page: https://zolkin1.github.io/generate-track-improve/

Zachary Olkin, William D. Compton, Aaron D. Ames

Humanoid LocomotionPerceptive Locomotion

VIDEOSep 22, 2026

GAM: 3D Grounding as a Robot Foundation

GAM uses pretrained 3D grounding to turn language, point, and box prompts into object-centric representations, improving manipulation robustness when scenes and targets move.

VIDEOSep 22, 2026

IR-SIM + Meta Habitat: Sim Robot Navigating a Video-Reconstructed Scene

Work in progress at the Auki lab in Hong Kong: Han Ruihua (HKU) IR-SIM is combined with Meta Habitat so a simulated robot navigates an environment previously reconstructed from video captured on another device.

VIDEOSep 21, 2026

Graph-as-Policy + Rust Executor: 5.42s vs 8.79s Median Pickup in YAM Sim

Weekend test in the YAM moving-cube sim: median pickup time is 5.42s with GaP plus a custom Rust executor versus 8.79s with native GaP, thanks to explicit decision nodes, dependencies and parallel execution.

VIDEOSep 21, 2026

Physical Agents SDK Turns Deployment Sites into Real2Sim Worlds

The Physical Agents SDK captures the robot's own cameras, sensors and LiDAR to turn deployment environments into ultrarealistic simulations; phones and tablets can contribute scenes, contributors are rewarded, and policies train where they will deploy.

PAPERSep 21, 2026

PredActor: Predictive Action Diffusion for Steerable Onboard Humanoid Control

Diffusion models provide a flexible framework for motion generation, but turning this flexibility into closed-loop humanoid control remains challenging. Hierarchical generator-tracker systems steer motion through reference trajectories, yet these references may exceed the capabilities of the downstream tracker, leaving physical feasibility and disturbance recovery largely to a separate control module. Action-only diffusion avoids this separation by directly generating executable actions, but provides no explicit future-state trajectory that can be steered toward test-time motion objectives. Joint state-action diffusion offers a natural alternative, but existing controllers often rely on privileged full-body states, while learned behavior selection and test-time motion steering remain only partially integrated. We present PredActor, a predictive action diffusion policy that unifies both steering modes in one directly executed policy using proprioception alone. Given proprioceptive history and optional task context, PredActor jointly predicts actions and an internal future-state trajectory that enables guidance: classifier-free guidance strengthens text-conditioned motion, while classifier guidance steers future states toward test-time objectives. Only actions are executed, requiring neither a motion-reference tracker nor privileged full-body states. In simulation, PredActor reaches 44 of 45 destination targets and achieves a text retrieval score of 0.539 versus 0.424 for conditional action diffusion, with similar disturbance survival. Rolling denoising and computation-preserving runtime optimizations reduce the complete callback to 16.790 ms median and 19.383 ms p95 on a Jetson Orin NX, within the 20 ms control period. Deployed on a Unitree G1, PredActor demonstrates text-conditioned motion, disturbance response, joystick control, and semantic interpolation in simulation and hardware.

Ye, Lei, Gao, Haibo, Li, Yitang

Humanoid ControlDiffusion policies

As an Amazon Associate, we earn from qualifying purchases.