Robotics & Physical AI

Tesla is redirecting people and factory resources toward its Optimus humanoid robot, but difficult hand assembly, unreliable sensing, limited autonomy, and worker resistance are complicating its scale-up.

Skild AI has demonstrated a humanoid robot that learns to dribble, shield the ball, tackle and recover from falls through repeated matches against earlier versions of itself. The project points to a possible way of reducing reliance on manually designed robot skills and reward functions.

A new startup founded by former Huawei large-model and multimodal AI leaders is pursuing a Large Physics Model, or LPM. Its central hypothesis is that robots may gain reusable, cross-task abilities through large-scale pretraining, much as language models did with text.

Spatial-Interactor moves spatial reasoning beyond static visual questions by training vision-language models to track state changes caused by actions, object motion, and viewpoint shifts.

High task success does not necessarily mean that an embodied agent understood its instruction. RoboFollow shows how low-entropy scenes allow policies to rely on vision and default behavior while largely ignoring language.

The $11 million XPRIZE Wildfire competition demonstrated that satellites, ground sensors, and drones can identify a blaze within minutes. Fully autonomous suppression remains much harder.

HuRo introduces a robotization pipeline that converts heterogeneous human manipulation videos into robot-aligned observations and action trajectories. Its experiments show that scaling this data can improve real-world manipulation and out-of-distribution generalization.

ShieldVLA brings Hamilton-Jacobi reachability into VLA safety fine-tuning, using visual observations to estimate whether a robot can still recover safely. The paper reports a 57% average reduction in cumulative safety cost across five navigation and manipulation benchmarks, alongside a 0.13 success-rate gain over SafeVLA.

CARE presents a data-centric framework for recovering robotic manipulation after execution drifts from the intended trajectory. It learns stage-specific failure patterns from real failed rollouts and uses them to generate corrective demonstrations and atomic recovery actions.

RoboDawn gives an agentic vision-language model a compact interface for translation, rotation, and gripper actions. Through closed-loop interaction and a small number of demonstrations, it transfers general-purpose reasoning to simulated and real robotic tasks.

A new study distills world-model representations into compact vision-language-action policies without putting future prediction inside the robot control loop. The approach improves manipulation results while preserving the baseline policy’s deployment cost.

Grounded Action Models (GAMs) make metric 3D grounding an explicit part of robot action prediction. By turning language, point, and box prompts into a shared object-centric representation, GAMs improve robustness when targets move, scenes change, or tasks require long-horizon control.

A new method adds a compact representation of interaction trends to 3D diffusion policies, allowing robots to use short observation histories to anticipate how an interaction is developing. It improves on DP3 without turning the policy into an explicit trajectory planner.

VABench evaluates more than object recognition: it measures whether multimodal models can acquire missing visual evidence, reason in a shared spatial frame, issue metric actions, and revise them after execution feedback. The results show a large gap between local spatial understanding and reliable embodied task completion.

Formerly known as UP.Labs, Vantora has raised $100 million and is shifting toward building physical AI startups exclusively for industrial customers. Its partners can invest in, use, and eventually absorb those ventures.

GPT-Policy combines a vision-language model with a context compiler and a constrained controller so robots can adapt to new tasks from demonstrations and feedback without gradient updates. Real-robot studies show that human videos can help, while action-aligned references are especially useful for contact-sensitive manipulation.

NATO-backed Scaleout Systems is putting lightweight computer-vision models on drones and forward-deployed hardware, allowing them to identify targets without a constant link to a central server. Its ALMA work has demonstrated autonomous detection, prioritization, geolocation, and attack-task execution.

Agility Robotics has unveiled Digit 5, a humanoid robot designed to work around people without relying on physical cages. It can reroute, stop, or squat when a nearby worker creates a potential collision risk.

PhysBrain 1.5 extends a vision-language model into a unified embodied foundation model that can interpret physical scenes, generate end-effector motion, and predict subsequent states. The paper reports an average score of 72.5 across 28 embodied understanding benchmarks for its 8B model.

Agent as Policy connects a general-purpose coding agent directly to a physical robot. It observes the scene, writes executable programs, issues motion commands, and revises its behavior using feedback from the real world.

ReactHuman evaluates whether embodied multimodal models can turn physical understanding into immediate action when household hazards suddenly occur. Its results show that reactive safety remains an unsolved capability.