🚀 Robotics Breakthroughs

Aquila Earth achieved a world-first by wirelessly powering a moving warehouse robot for 24 hours using infrared laser beams, setting records for highest total laser power transferred and longest continuous duration. This demonstration validates power-beaming as a viable solution for continuous autonomous operations, potentially eliminating battery downtime in logistics and other industries.

A humanoid robot completed a half-marathon faster than any human, demonstrating unprecedented endurance and locomotion control. This milestone suggests rapid progress in bipedal robotics and could accelerate adoption in logistics and emergency response.

NVIDIA Research's Eureka AI agent autonomously generates reward algorithms for robot training, outperforming human-written algorithms on over 80% of complex tasks. This represents a significant advancement in autonomous skill acquisition for robotics, potentially accelerating the development of general-purpose robots.

McGill engineers created ultra-thin graphene oxide sheets that fold like origami, enabling soft robots to walk, twist, and sense motion. This addresses the key scalability challenge in soft robotics by combining actuation and sensing in a single material.

Figure AI's F.03 humanoid robot autonomously sorted nearly 250,000 packages over 200 hours in a warehouse stress test, achieving near-human speed. This marks a significant milestone in deploying general-purpose humanoid robots for real-world logistics automation, demonstrating reliability and scalability.

NVIDIA Research introduced NeRD, Dexplore, and VT-Refine at CoRL 2025, enhancing simulation with learned dynamics, enabling efficient exploration, and combining vision-tactile sensing for bimanual assembly. These innovations represent a first-of-its-kind capability in integrating multiple neural approaches for practical robot learning.

This paper presents the first experimental demonstration of non-line-of-sight (NLOS) imaging applied to autonomous navigation, using single-photon LiDAR to detect hidden objects. It enables robots to perceive and avoid obstacles around corners, significantly expanding perception capabilities beyond line-of-sight.

π0.7 is a new AI model that allows robots to generalize to tasks they were never trained on, using plain-language instructions. This represents an early but significant step toward a general-purpose robot brain, moving beyond narrow, pre-programmed behaviors.

AgiBot's A2 humanoid robot completed a 65-mile cross-province trek from Suzhou to Shanghai, setting a Guinness World Record for the longest journey walked by a humanoid robot. This demonstrates significant progress in bipedal locomotion endurance and real-world navigation for humanoid robots.

Unitree Robotics has introduced the GD01, the world's first production-ready manned mecha, capable of transforming between bipedal and quadrupedal modes. Priced at $650,000, this 500 kg robot represents a major step from concept to commercial reality in humanoid robotics.

Ouster's Rev8 lidar captures 3D spatial and color data simultaneously in a single sensor, eliminating the need for separate cameras. This simplifies sensor fusion, reduces cost, and enhances perception for autonomous systems. It is the first mass-produced native color lidar, marking a significant industry milestone.

Sony AI's Project Ace is the first autonomous robot to defeat elite human players in competitive table tennis, showcasing physical AI that integrates perception, decision-making, and action faster than humans. This milestone demonstrates significant progress in real-world robotics and AI, bridging the gap between simulation and physical interaction.

Generalist AI's GEN-1 is the first general-purpose embodied foundation model to achieve mastery-level performance on simple physical tasks, crossing a new threshold in robotics. This represents a significant step toward versatile AI systems that can operate in the real world, moving beyond narrow, task-specific robots.

Figure AI increased production throughput by 24x in under 120 days, reaching a rate of one Figure 03 humanoid robot per hour with over 80% yield. This marks a first-of-its-kind manufacturing milestone for humanoid robots, demonstrating rapid scaling from lab to factory.

Fujitsu's new spatial world model technology enables AI to predict behaviors of actors and objects in space, improving behavioral intention estimation accuracy by 3x. This represents a significant step in human-robot collaboration, demonstrated at CES 2026.

Deep TAMER allows robots to learn tasks via real-time human critique, combining deep learning with interactive training. This approach reduces the need for large datasets and enables faster, more intuitive skill acquisition.

Columbia Engineering researchers developed a robot that autonomously learns realistic lip movements for speech and singing by observing its own reflection and human videos, without explicit programming. This self-supervised learning approach significantly reduces the uncanny valley effect in human-robot interaction.

Georgia Tech's EgoMimic algorithm allows humanoid robots to learn complex manipulation skills directly from first-person human videos without manual programming. This approach significantly reduces the data and training required for robots to perform household tasks like laundry and dishwashing, marking a major step toward practical home assistance.