Robotics & Physical AI

MobileVLA-R1 2.0 combines structured embodied reasoning, reinforcement learning, and a reasoning-conditioned action decoder to narrow the gap between language understanding and reliable robot execution. The framework is evaluated across navigation, quadruped control, and humanoid mobile manipulation.

Robot foundation models can perform well in familiar settings while relying on visual cues that are irrelevant to the task. Latent Interface Training, or LIT, separates goal-directed action learning from visual conditioning and uses pose supervision to improve robustness under distribution shifts.

A roughly $4,000 Unitree robot dog shows how far the cost of legged robotics has fallen. It also reveals the practical limits that still separate an impressive machine from a useful household tool.

MaP-WAM treats long-term memory as planning evidence rather than repeatedly feeding the full history to a robot policy. It converts episodic multimodal context into segment-level plans with visual guidance and adaptive execution.

Show-Harness introduces a compact semantic interface that connects a vision-language model’s intent to robot actions without requiring embodiment-specific pretraining. Its GUMI extension also enables demonstration collection through a graphical interface rather than dedicated teleoperation hardware.

Maven Robotics is emerging from stealth with a $100 million Series A and robots already operating in logistics facilities. Rather than lead with humanoids or foundation models, it is targeting concrete warehouse workflows first.

DriveZero combines closed-loop reinforcement learning with a vision foundation model to reduce autonomous driving’s dependence on human trajectory logs. Its design separates perception from action, then reunifies them through distillation into a camera-only planner.

RoboTok retrieves human manipulation videos that resemble a given demonstration, using hand-motion similarity rather than relying mainly on visual appearance. Its actor-centered 3D hand-trajectory representation is designed to make retrieval more robust across viewpoints, occlusions, and changing scenes.

Qwen-Drive-1.0 explores a unified framework that combines 3D perception, visual question answering, and motion planning while retaining the backbone of a pretrained vision-language model. A BEV perception head and a planning expert connect shared representations to inspectable scene structure and future vehicle trajectories.

Nori Robotics has introduced a low-cost bimanual mobile platform for robotics developers and researchers, priced at $1,688 and scheduled to ship in fall 2026. Its design trades legged mobility and onboard compute for affordability, repairability, and easier experimentation.

NavMCP combines the task-level reasoning of a vision-language model with the closed-loop execution of a navigation foundation model. The result is a persistent exploration loop for physical-world agents.

Robot trajectories are expensive and difficult to scale, making representation quality as important as data volume for general-purpose vision-language-action models. VLAct uses multi-embodiment continued pre-training to improve transfer under constrained data and compute.

Washington is tightening access for foreign-made drones and advanced robots, but the measures may not produce a clean U.S.-China split. China’s manufacturing scale and cost advantage could redirect competition toward other regions.

Meta is testing robots that can reset servers, manipulate cables, inspect equipment, and move heavy racks inside its data centers. The effort could ease labor shortages while putting pressure on some technician roles.

Zero-WAM treats a human manipulation video as an in-context task specification for a robot, rather than requiring parameter updates for every new task. The project introduces the HumanGen dataset and reports a 47.0% average success rate on seven unseen RoboTwin 2.0 tasks.

TacForcing is a streaming action-generation framework that lets vision-language-action policies use tactile feedback as actions are being executed. It replaces a separate high-frequency reactive controller with a streaming action expert and reports stronger average success on simulated and real contact-rich tasks.

Anthropic has introduced a research preview of its Model Hardware Standard, or MHS, a proposed common interface for AI agents to discover and operate robots, microscopes, cameras, and laboratory instruments. Rather than giving Claude one permanent body, the approach lets it use many connected devices as temporary extensions.

StreamPI adds temporal reasoning to single-frame vision-language-action models without introducing extra parameters. Its instruction-anchored attention and randomized interval training are designed to improve memory, spatial perception, and robustness to asynchronous robot inputs.

GigaBrain-0.7 proposes a three-system architecture that connects visual-language understanding, world prediction, and robot action. Trained on more than 37,000 hours of heterogeneous embodied data, it targets stronger transfer across tasks and robot embodiments.

WorldToken combines multiview images, proprioception, and task conditioning into one token for each policy timestep, then models the resulting sequence causally. Its experiments point to the importance of target-domain data and long temporal context in robotic imitation learning.

PhysCaP extends code-as-policy robotics with physics-informed active perception. Instead of relying only on what a camera can see, the agent uses targeted interactions and proprioceptive feedback to infer hidden properties such as mass and stiffness.