Physical AI News (April 10, 2026 Issue)
Update Date: 2026/4/10
Executive Summary
Physical AI on April 9, 2026, saw advancements in intelligent element technologies across both software and hardware, including TAMEn's tactile manipulation learning, AgiBot GO-2's integrated reasoning and action, Genie Sim 3.0's Spatial World Model, Tesla FSD's MLIR overhaul, SiMa.ai's memory-centric edge semiconductors, FANUC's digital twin integration, RoboSense's surge in robotics LiDAR, and RoSHI's outdoor human data collection. The common themes are a shift from vision-centric to multi-modal approaches including touch and embodiment, closed-loop design that learns even from failure states, world models premised on Sim2Real, and computational resource optimization for immediate implementation. This indicates that the competitive axis of physical AI is shifting from model scale to reproducibility and mass-production deployment capability that works stably in the field.
※ The content of the created article was input into Gamma to automatically generate slides. If you find slides easier to view, please take a look here.
1️⃣ TAMEn: Tactile x Closed-Loop Data Collection Engine More Than Doubles Manipulation AI Accuracy
📎 Source: arXiv:2604.07335
The success rate of dual-arm manipulation, which was 34% on average with vision alone, improved in three stages: to 55% with the introduction of tactile feedback, to 65% with pre-training using large-scale single-arm tactile data, and to 75% by adding 10% of online recovery data near failure states. Targeting contact-dominant tasks such as cable routing and dishwashing, it implements wearable dual-mode collection (high-precision MoCap + portable VR) and a closed-loop system including feasibility checks. This paper quantitatively proves that 'multi-modal contact information and learning from failure states' are more effective for improving physical AI performance than 'model scaling'.
2️⃣ AgiBot 'Genie Operator-2 (GO-2)' — VLA Foundation Model Integrating Reasoning and Action Released
📎 Source: AGIBOT Official
China's AgiBot has released GO-2, a VLA foundation model that bridges the 'Semantic-Actuation Gap' between reasoning (planning) and action execution. It adopts Action Chain-of-Thought + an asynchronous dual system (low-frequency semantic planning x high-frequency motion tracking). It recorded a LIBERO average success rate of 98.5% and an 82.9% success rate in Sim2Real real-machine tests in Genie Sim 3.0. It claims ~10x training efficiency, 2-4x success rate, and 50%+ reduction in data requirements. Acceptance to CVPR/ACL 2026 was also announced.
3️⃣ AgiBot 'Genie Sim 3.0' — Upgraded to a Spatial World Model that Generates 3D Worlds with Natural Language
📎 Source: AGIBOT Official / GitHub
An update that integrates simulation -> data -> evaluation -> RL, centered on the 'Spatial World Model' that generates interactive 3D environments from text/image input. It has prepared 200+ tasks/100,000+ scenarios/10,000+ hours of synthetic data, aiming for a discrepancy of <10% between simulation and real-world tests. It supports massively parallel execution of RL with 1000Hz physics simulation (decoupling rendering and physics).
4️⃣ Tesla FSD v14.3 — 20% Improvement in Response Speed with Complete MLIR Compiler Overhaul
📎 Source: Tesla Oracle
Tesla has begun distributing FSD v14.3 (firmware 2026.2.9.6). By rebuilding the MLIR-based AI compiler and runtime from scratch, it achieved a 20% improvement in response speed and improved model iteration speed. The neural network's vision encoder has also been strengthened, improving 3D geometric understanding of complex intersections, visibility in low-light scenes, and traffic sign recognition accuracy. RL training for edge cases, such as avoiding small animals and responding to vehicles violating right-hand traffic, has also been enhanced. This update simultaneously raises the 'judgment accuracy and safety of autonomous driving' through a combination of compiler optimization and model improvements.
5️⃣ SiMa.ai x Micron — Strategic Investment in Memory-Centric Physical AI
📎 Source: PR Newswire
Edge AI chip company SiMa.ai has secured a strategic investment from Micron Technology. By integrating Micron's LPDDR5X memory into the Modalix MLSoC platform, it has strengthened the production system for physical AI chips that achieve both high bandwidth and ultra-low power consumption. For robots, autonomous vehicles, and industrial automation, it provides a foundation for 'perception -> reasoning -> action' that completes LLM/VLM inference on the edge without relying on the cloud. It is available immediately in SoM (System-on-Module) format, allowing for seamless integration into existing platforms.
6️⃣ FANUC x NVIDIA — Physical AI Fully Integrated into Over 2 Million Industrial Robots
📎 Source: National Today
Industrial robot giant FANUC is deepening its partnership with NVIDIA. By integrating Jetson edge modules and Omniverse/Isaac Sim into its robot portfolio and simulation software 'ROBOGUIDE', it reduces commissioning time and costs through pre-verification with digital twins. It enables non-programming robot operation by automatically generating Python code from voice commands. Through physical AI, it bridges the 'gap between virtual design and real-world production', accelerating the deployment of adaptive automation to manufacturing sites suffering from labor shortages.
7️⃣ RoboSense Q1 2026 Results — LiDAR for Robots Up 1,458% Year-on-Year, Surpassing Automotive
📎 Source: PR Newswire
LiDAR leader RoboSense (2498.HK) announced its sales results for Q1 2026. Out of a total of 330,300 units (up 204.1% year-on-year), the robotics and other segments reached 185,500 units, a 1,458.8% increase compared to the same period last year. For the first time in the company's history, robotics applications surpassed automotive ADAS (144,800 units). It holds the top market share in five fields: autonomous lawn mowers, delivery robots, humanoids, and commercial cleaning robots. By mass-producing fully digital LiDAR using its proprietary SPAD-SoC/VCSEL chips, it is leading both the automotive and robotics markets.
8️⃣ RoSHI: Connecting Outdoor Human Data to Humanoid Learning with Low-Cost IMU x SLAM
📎 Source: arXiv:2604.07331
A wearable system that fuses a group of low-cost IMUs (9 units, approx. 350 USD) + SLAM from Meta's research AR headset 'Project Aria'. It acquires long-term, occlusion-resistant 3D full-body posture data in outdoor and daily environments outside the laboratory, connecting it to humanoid behavior learning. By complementarily fusing IMUs (strong against occlusion and high-speed motion) and SLAM (long-distance global consistency), it simultaneously achieves 'portability, occlusion resistance, and global consistency'. This paper also demonstrates its application to RL policy learning on a Unitree G1 real machine.
Comprehensive Analysis
The key feature of the topics on April 9, 2026, is that the performance of physical AI is beginning to be determined not by the intelligence of a single model, but by the degree of integration of peripheral stacks such as contact information, failure recovery, world generation, compilers, memory, sensors, and digital twins. TAMEn and RoSHI demonstrate "how to acquire real-world physical data cheaply, for longer periods, and in large quantities," while GO-2 and Genie Sim 3.0 show "how to translate semantic understanding into actionable behavior." Meanwhile, Tesla, SiMa.ai, Fanuc, and RoboSense are competing on "whether that intelligence can be mass-deployed with low latency and safety." In other words, the competition is shifting from comparing research accuracy to building closed industrial systems that encompass data collection, learning, evaluation, inference, implementation, and supply chains, and the next winners are likely to be the groups of companies that can bundle all these processes together.
Future Points of Interest
With the value of tactile and near-failure data quantified by TAMEn, the real difference in manipulation AI will no longer be the competition for visual data volume, but rather how contact logs and recovery trajectories can be standardized and continuously collected on actual machines.
As GO-2 style asynchronous dual-layer control spreads, the evaluation axis for VLA will shift from benchmark success rates to a competition in control architecture, specifically whether planning update cycles and motion tracking cycles can be designed separately.
Looking at the trends of Genie Sim 3.0 and Fanuc, the future battleground is likely to shift to operational Sim2Real, where the focus is not on "how realistic the simulator is," but on "how quickly and continuously real-world discrepancies can be absorbed."
The moves by Tesla and SiMa.ai indicate that in physical AI, algorithmic improvement alone is insufficient, and that optimization of the computing infrastructure—including compilers, memory bandwidth, and power consumption—directly dictates safety and responsiveness.
The change in RoboSense's sales composition is a sign that the main battlefield in the sensor market is expanding from ADAS alone to robots, and competition for perception components that can be mass-produced for cleaning, delivery, lawn mowing, and humanoid applications is likely to accelerate.
As low-cost wearables like RoSHI become widespread, humanoid learning will move away from dependence on expensive laboratory measurements and enter a phase of capturing long-term behavioral data in everyday environments, significantly broadening the diversity of behaviors.
