MolmoBOT uses large-scale simulation to generate diverse data, aiming to enable zero-shot manipulation policies without requiring real-world robot data. The system includes the MolmoBot-Engine for procedural data generation and MolmoBot-Data, a dataset of 1.7 million expert trajectories. Three policy classes, including MolmoBot, MolmoBot-SPOC, and MolmoBot-Pi0, were trained on this synthetic data. These policies achieved 79.2% success rates in real-world pick-and-place tasks with zero real-world fine-tuning. The work demonstrates that extensive simulated data can facilitate effective sim-to-real transfer.
Warhound raised $200 million to genetically engineer war dogs for military and law enforcement tasks. The project aims to enhance dogs' health, longevity, and sensory abilities through genetic editing. Anduril will provide associated hardware and software, such as protective gear for situational awareness. The company plans to use cell lines from breeds like the Belgian Malinois and German Shepherd to create these enhanced dogs. The initial goal is to develop these dogs, with the first batch expected within two years, and seek government approval.
AI agents like Hermes and Nemotron plan an obstacle course, which is then used to create 3D assets via Astra for simulation. These assets are assembled in NVIDIA Isaac Sim, incorporating Newton for physics modeling and the Omniverse libraries for rendering and sensor simulation. A perception and locomotion policy is then trained at scale in NVIDIA Isaac Lab using parallel environments and observations. The trained policy is validated in Isaac Sim before being prepared for sim-to-real transfer on physical robots.
A quadruped robot autonomously solved a giant Rubik’s Cube using only its feet. Researchers developed the Reflex framework to enable humanoids to perform visual reactive whole-body control. Other advancements include multi-fingered robot hands with tendon-driven actuation and biohybrid robots with living tissue and soft scaffolds. Progress in robotics is also shown in motion tracking, dual robot attachments, and mobile systems for lunar exploration. The humanoid form factor is being justified in robotics through engineering and business perspectives.
The 140th Canton Fair launched a dedicated zone for service robots as part of its "Advanced Manufacturing" program. International buyers shifted focus from scripted demonstrations to assessing the real-world reliability of robotic systems. The robot zone featured live operational demonstrations of cleaning-and-delivery robots, humanoid platforms, and inspection robots. Buyers prioritized whether robotic systems could consistently handle full-shift commercial and industrial workloads. Sixty-four percent of participating enterprises deployed industrial-internet-enabled or AI-powered solutions.
NVIDIA expanded its Isaac platform with training tools, datasets, and a runtime framework alongside Agile Robots' Agile ONE SN humanoid platform. This new platform integrates NVIDIA Isaac GR00T with Jetson Thor for onboard computing. The Agile ONE SN features a 40 kg humanoid body with dexterous hands, sensors, and the GR00T integration for coordinated movement. NVIDIA provides open-source tools like Isaac Lab 3.0 and the Isaac OS to facilitate training and real-time operation. This ecosystem is attracting research interest from institutions like Stanford and ETH Zurich.
HONOR A1 is a compact humanoid robot developed by HONOR, measuring 136.9 cm tall and weighing 40 kg. Its demonstrated capabilities focus on physical mobility, including obstacle avoidance, running 100m, and walking stairs, with a maximum speed of 10 km/h. The robot's skill scores indicate its navigation is developing, while manipulation capability is rated at zero. A1 is currently positioned as a prototype with a listed price of USD 25,000, emphasizing locomotion over object handling. The development showcases HONOR's extension of smartphone engineering into humanoid hardware.
Generative Bionics plans to build a 5,000 square meter humanoid training facility in Italy by the fourth quarter of 2027. This Physical AI Data Factory will reproduce industrial environments to train and validate the GENE.01 humanoid before customer deployment. The facility will collect physical interaction data, including vision, motion, and multimodal skin signals, to train world action models. The company will create digital twins for each physical environment and robot configuration to simulate thousands of variations. AMD's Helios will provide the computing foundation for processing this physical and synthetic data.
RH56F1 is a robotic hand from Inspire Robotics, noted for its developing performance rating of 5 out of 10. It has a Degrees of Freedom rating of 2 out of 5 and a Strength rating of 3 out of 5, indicating a stronger rating in strength. The hand measures 78 × 165 mm and weighs 0.64 kg, providing physical specifications. Priced at USD 8,500, it is commercially available for teams to evaluate hardware. The separate category scores offer a nuanced comparison beyond the overall score.
The G1 D is a wheeled humanoid introduced in November 2025, designed for AI training data collection, research, and education. It features lifting columns instead of legs and seven degrees of freedom in each arm, totaling 59 degrees of freedom. Supported capabilities include sorting goods, product assembly, and household handling, focusing on manipulation tasks. The model is available in a production configuration at a price of USD 51,900. Unitree Robotics positions the platform as a manipulation base, with mobility being a secondary feature.
The Linker Hand L6 is a robotic hand from LinkerBot, designed with a compact mechanical structure emphasizing dexterity and grip strength. It has a Humanoid.Guide Hand Score of 6 out of 10, indicating a capable level of performance. The hand weighs 607 grams and utilizes a rigid linkage transmission powered by a brushless motor and ball screw linear drive. The equal ratings for Degrees of Freedom and Strength are 3 out of 5, signifying a balanced assessment rather than exceptional performance. The hand is listed as in production with an advertised price of USD 2,800.
Unitree Robotics' H2 D is a wheeled humanoid designed for research, training data collection, and manipulation tasks. It features a mobile base, a height-adjustable column, and two seven-degrees-of-freedom arms, with a listed height of 170 cm and a weight of 120 kg. The platform can perform tasks like obstacle avoidance, sorting goods, and product assembly using its sensing system and articulated hands. It is supported by a Linux-based system with ROS 2, targeting researchers and developers for focused experimentation. The robot is currently listed in production at a price of USD 55,000.