Physical AI News (May 29, 2026 Issue)
Update Date: 2026/5/29
Executive Summary
May 28, 2026, clearly showed the trend of Physical AI moving from research results to implementation and market expansion. RoboSense's LiDAR shipments for robotics surpassed those for ADAS for the first time, indicating that the primary demand for sensors has expanded from automobiles to robotics in general. In terms of research, the omnidirectional mobile robot Argus, tactile CoP representation, surgical assistance imitation learning, magnetically driven microrobot VLA, Unitree G1 high-speed running, and large-area robot skin have emerged. Furthermore, from the acquisition of FORT Robotics and FANUC's industrial demonstrations, the transition to the actual operation phase, including safety supervision, remote intervention, and factory implementation, is progressing.
*The created article content was input into Gamma to automatically create slides. If you find slides easier to view, please take a look here.
1️⃣ RoboSense: LiDAR shipments for robotics surpass automotive ADAS
📎 Source: RoboSense / PR Newswire / Gasgoo
RoboSense announced its Q1 2026 financial results. LiDAR shipments for robotics reached 185,500 units, a 1,458.8% increase year-on-year, surpassing shipments for ADAS for the first time in the company's history. Total shipments were 330,300 units (a 204.1% increase year-on-year), with robotics accounting for approximately 56% of the total. Revenue was 458.8 million yuan (a 39.9% increase year-on-year), and the ADAS order backlog exceeds 9 million units. Adoption is accelerating in a wide range of robot markets, including humanoids, delivery, cleaning, and lawn-mowing robots, making this an important financial report showing that the growth axis is structurally shifting from automobiles alone to robotics in general.
2️⃣ Argus: A 20-legged 'no front or back' robot and the proposal of dynamic isotropy
📎 Source: EurekAlert / Duke University / Science Robotics DOI
A research team at Duke University announced the robot 'Argus,' which features 20 telescopic legs and depth cameras arranged radially, in Science Robotics. They explored over 1,500 shape candidates and evaluated acceleration and movement performance in all directions using an index called 'Dynamic Isotropy.' While common quadrupeds, humanoids, and drones score below 0.6, Argus recorded 0.91. It demonstrated movement in forests, sandy ground, wet ground, wall climbing, and even when three legs were damaged, showing the potential for function-first robot design principles that differ from the humanoid-centric approach.
3️⃣ FORT Robotics × Mapless AI: A trust foundation integrating remote supervision and active safety
📎 Source: FORT Robotics
FORT Robotics has acquired Mapless AI, which possesses vehicle teleoperation and autonomous supervision technology. With this, FORT's Trust Platform will add remote Human-in-the-Loop teleoperation and onboard active safety features. The company explains that it is expanding from conventional safety-certified machine control to an autonomous operation architecture under supervision. It is attracting attention as a foundation for supporting safety stops, human intervention, and division of responsibility when scaling autonomous machines in real environments such as construction, logistics, defense, and last-mile delivery.
4️⃣ Beyond Binary: Achieving dexterous Sim2Real manipulation with CoP tactile representation
📎 Source: arXiv:2605.28812 / Project Page
A research team from ETH Zurich, UC Berkeley, and others proposed the Center-of-Pressure (CoP) representation, which compresses tactile information as physical quantities. CoP retains contact points and force vectors, mapping them to real-machine tactile sensations through sensor calibration using differentiable dynamics. In peg insertion and ball balancing with 3-finger and multi-finger hands, CoP-conditioned policies achieved zero-shot Sim2Real transfer, outperforming binary contact or raw taxel inputs. This is a result that advances tactile-based robot learning in contact operations where visual reliance is difficult.
5️⃣ Open Surgery IL: π0 completes 92% of stitches with surgical assistance robot
📎 Source: arXiv:2605.28736
Research evaluating general-purpose imitation learning for suture assistance during surgery has been published. The research team collected 160 remote operation demonstrations with 32,374 frames using an open-source robot arm and compared ACT, Diffusion Policy, SmolVLA, and π0. Under ideal conditions, the four methods showed a success rate of 50-75%, with π0, which has a pre-trained VLA backbone, showing the highest data efficiency and smooth trajectories. In surgeon/robot suturing tests, π0 achieved a 92% stitch completion rate, and depth estimation and end-effector design emerged as the next challenges.
6️⃣ Mag-VLA: Two-arm VLA control for magnetically driven microrobots
📎 Source: arXiv:2605.28486
Mag-VLA is a VLA model that operates magnetically driven microrobots using two robot arms equipped with magnets. It adapts Qwen2.5-VL-7B with LoRA to generate actions from visual observations and language instructions. By using a phase classifier to detect task progress and a phase-conditioned ACT decoder, it achieved chronologically consistent two-arm control. In real-machine experiments, it recorded a 90% arrival success rate for all tasks, and 80%, 70%, and 50% transport success rates by difficulty level. It showed that VLA can also be applied to non-contact operations at a micro-scale.
7️⃣ SPRINT: Humanoid sprinting up to 6m/s with Unitree G1
📎 Source: arXiv:2605.28549
SPRINT is research that proposes frequency-adaptive spectral motion priors to achieve high-speed humanoid running. It expresses the periodicity of human running in the frequency domain and generates joint trajectories that support a wide speed range from just five types of reference motion sequences. Field experiments on the Unitree G1 succeeded in zero-shot Sim2Real transfer, achieving a sprint speed of up to 6m/s. It maintains gait transitions and biomimetic naturalness even at high speeds, providing significant implications for improving data efficiency in humanoid movement control.
8️⃣ EIT-Pneumatic Robotic Skin: Improving force reconstruction accuracy for large-area robot skin
📎 Source: arXiv:2605.28468
A hybrid robotic skin combining EIT and pneumatic tactile sensing has been announced. The sensors are fabricated using 3D printing and spray coating, aiming for stable force distribution estimation over large areas through inverse problem reconstruction with Tikhonov regularization and per-pad pneumatic calibration. In indentation experiments using load cells, sensitivity non-uniformity was improved compared to the EIT-only baseline, with the coefficient of variation decreasing from 0.31 to 0.14. Integration into a humanoid chest also demonstrated reliability across diverse contacts, including multiple simultaneous contacts.
9️⃣ FANUC America: Announces Physical AI demos for industry at Automate 2026
📎 Source: FANUC America / PR Newswire / FANUC Automate 2026
FANUC America has announced its exhibition details for Automate 2026, to be held at McCormick Place in Chicago from June 22 to 26. At the booth, they will conduct a bolt-tightening demo on a moving engine block using a CRX-20iA/L collaborative robot, showcasing dynamic tracking powered by Inbolt Physical AI and NVIDIA Jetson processing. They also plan to demonstrate practical factory-floor Physical AI applications by integrating advanced features—including "CRX Vibe Coding" (which converts natural language instructions into AI-generated Python code for robot motion), 3D vision, digital twins, and ROS 2 connectivity—into existing industrial robots.
Comprehensive Analysis
The key takeaway from the topics on May 28, 2026, is that Physical AI is shifting from a "performance competition for individual humanoid robots" to a comprehensive system competition that includes sensors, tactile feedback, safety infrastructure, remote supervision, and industrial integration. The structural change in LiDAR shipments indicates an expansion in the robot market, covering delivery, cleaning, lawn mowing, and humanoids. Meanwhile, research into Argus and robotic skin highlights the importance of body design and contact understanding beyond mere human imitation. While VLA and imitation learning are expanding into surgery, micro-manipulation, and factory tasks, safety stops, human intervention, and reliability assessment are essential in real-world environments, shifting the focus of technology from "moving AI" to "trusted AI."
Future Points of Interest
With robotics-focused LiDAR surpassing ADAS, the growth metrics for sensor companies are entering a phase where they will be evaluated not just on automotive adoption numbers, but on their ability to scale horizontally across different robot types.
The dynamic isotropy of Argus is an important design evaluation axis that is not biased toward humanoids or quadrupeds; for disaster response and off-road exploration, a "body that can move even when damaged" may be superior to "naturalness."
Tactile CoP representation and EIT pneumatic robotic skin are attracting attention as foundational technologies that complement vision-centric robot learning and enhance the reproducibility and safety of contact-based tasks.
The application of VLA to surgical assistance and magnetically driven microrobots shows signs that Physical AI is expanding beyond large robots into the medical and micro-manipulation domains.
The moves by FORT and FANUC clarify that what is important in practical operation is not just model performance, but also remote supervision, division of responsibility, and integration with existing equipment.
