Gemini Robotics' latest release introduces walking as part of the action space, no longer a separate subsystem. This innovation, achieved through policy conditioned on language instruction, represents a significant advancement over previous models with fixed bases and lack of legs. The demo showcasing this capability is intentionally mundane to highlight the complexity behind the technology.
1X's Neo robot features soft, helpful hands with actuators designed to replicate human hand movement, providing 25 degrees of freedom and enabling tasks like opening doors, lifting heavy objects, and plugging itself in. The hands have an IP68 waterproof rating and can grip odd shapes, detect slipping, and move extremely quickly. The robot's design is inspired by characters like Baymax, with a focus on being soft, supple, and safe for human interaction. Early access pricing starts at $20,000 or $500 per month, and the robot is currently partly teleoperated, allowing humans to remotely control it in "Expert Mode."
NVIDIA Alpamayo 2 Super is now available for commercial use, offering advanced reasoning capabilities for robotaxis and autonomous vehicles. Built on NVIDIA Cosmos 3 Super Reasoner and post-trained with reinforcement learning, the model advances the AV ecosystem with leading multitask capabilities and open commercial licensing. The model ranks first on LingoQA and other benchmarks, demonstrating state-of-the-art reasoning for driving-centric scenarios.
World Action Models (WAMs) build on video world models instead of vision-language models (VLMs), enabling robot policies with stronger physical generalization. A strong foundation for post-training WAMs is the open NVIDIA Cosmos 3 model, which provides practical deployment at multiple tiers and measurable improvements in policy performance. Unlike VLA models, WAMs learn from diverse data and generalize in the open world, adapting to new robots with few demonstrations.
Persona AI has deployed humanoids to South Korea for early integration work ahead of its Gen 3 deliveries at the end of the year. The deployment follows a successful teleoperated welding demo by Persona AI's Gen 1 robot in July, which validated its core actuation and kinematics. The company is partnering with HD Hyundai to commercialize humanoid labor for shipyard environments, and the arrival of hardware in Korea aligns closely with this partnership.
Sunday Robotics introduced ACT-2, a model achieving a 99.1% autonomous success rate in folding laundry across unseen home environments without requiring per-home data collection or expert demonstrations. The system closes the "generalization gap" by scaling pretraining on diverse data and exhibits emergent behaviors such as recovering dropped garments and continuing tasks despite physical disturbances. A new evaluation framework called a "Solve" is proposed, which evaluates tasks based on performance, scope, and adaptation cost rather than isolated demonstrations.
Regular price ROSCon Global tickets can be purchased until Tuesday, August 24th, 2026 at 6:59 AM UTC.
Autonomous robots are transforming polar science in the Arctic and Antarctic by collecting data from remote and inhospitable environments. They operate underwater, on the ocean surface, and in the air, allowing scientists to monitor glaciers, oceans, wildlife, and climate processes over long periods. Autonomous vehicles, such as gliders and drones, can work in places humans cannot, reducing risks and increasing efficiency. Researchers believe artificial intelligence will become increasingly important in coordinating different sources of information, but robots are not expected to replace scientists entirely.
Yusen Logistics Americas, a global supply chain logistics company, has deployed Destro AI's warehouse coordination platform as part of its strategic collaboration with the provider. The deployment aims to improve operational efficiency, increase execution consistency, and deliver value to customers through real-time coordination across people, equipment, and material movement. The platform continuously analyzes warehouse conditions, optimizes work assignments for employees and autonomous mobile robots, and provides real-time operational visibility. The initial focus is on optimizing cart movements in transload operations, with plans to evaluate additional automation capabilities in the future.
A Qualcomm-powered robot called 4NE-1 collapsed spectacularly on stage during a press event at Computex 2026, just seconds into its presentation. The robot's slow shambling walking style was mocked by some, but the real moment of comedy came when it suddenly and noisily crashed to the ground, prompting stagehands to quickly cloak and carry it away. The presenter continued his speech as if nothing had happened, and the audience applauded politely for him, rather than the robot's mishap. The incident has been compared to a Windows 98 software crash event, which may have led to the creation of the infamous "USB Cart of Death".
Zoox has unveiled the final production design of its robotaxi, featuring a new color scheme and improved rider comfort features such as screens, controls, and drink holders. The vehicle's cabin has been designed to be "carriage-like" with no steering wheel or pedals, seating four adult passengers facing each other. The updated materials are built to withstand urban wear and tear while offering higher contrast for easier spotting of personal items. Zoox plans to produce 100 vehicles per week at its new manufacturing facility in Hayward, California, as it prepares for full approval from the NHTSA.
Laboratories can only simulate certain aspects of real-world environments, making it challenging for robots to be resilient and predictable in deployment. Early field testing helps identify flaws early in design, speeding up the design feedback loop cycle. Building a strong robotics deployment pipeline requires iterating development on platforms and designing technology that people will trust, including factors like signaling intentions and safety points. The role of the feedback loop in deployment involves using real data to optimize performance and strengthen design autonomy through iteration-based innovation.
LLP, a single deployable policy, enables autonomous perceptive whole-body locomotion on a humanoid in outdoor settings using only onboard sensing. The system acquires terrain-conditioned skills from sparse seeds and learns autonomous skill transitions from reward. LLP outperforms existing humanoid systems by tracking velocity in open terrain, executing whole-body traversal at obstacles, and resuming locomotion afterward, with no reference input or skill label. Simulation and real-world experiments show high success across benchmarked terrains and unseen obstacle variations.
Semantic haptic feedback enhances dexterous robotic teleoperation by using abstract patterns to convey critical information about robot states, simplifying hardware requirements and enabling one-to-many mappings. The approach categorizes robot states into "Confirmations" and "Exceptions", implements a modular haptic rendering pipeline in robot simulation, and delivers semantic haptic feedback through pneumatic and vibrotactile wristbands. Three evaluation studies show that semantic haptics performs similarly to other feedback approaches in unimanual tasks but achieves superior performance in bimanual tasks with reduced task workload, increased situational awareness, and overall preference.
Industrial humanoid robots face a lack of functional safety certification due to the fail-passive gap, where actively-controlled states on the robot-side violate the assume-safe state assumption. A certified external safety chain is used to locate this issue precisely, pinpointing it to the robot-side reaction chain. The approach involves analyzing the balancing policy and its residual risk, but hosting an industrial software-defined automation controller reinforces the fail-passive gap and localizes it to the SDA-to-balancing-policy interface.
Biconvex optimization is presented for minimum-time motion planning around convex obstacles, ensuring convergence and supporting derivative constraints to arbitrary order. The method jointly convexifies the objective and constraints through a change of variables, handling collision avoidance via time-varying separating planes. The approach reliably produces high-quality trajectories with computation times comparable to state-of-the-art decomposition-based motion planners in drone navigation and dual-arm bin unloading experiments.
Researchers have proposed a method to control barrier functions via Minkowski operations for safe navigation among polytopes in robotics. The approach computes an exact Signed Distance Function (SDF) using companion convex programs and leverages geometric properties of 2D Minkowski operations to derive a unified analytical expression for the gradient of the SDF.
Researchers Mohammadali Rashidioun, Michael Sosa, and Petras Swissler propose a novel contact-driven localization method for modular robots that uses local communication through binary contact information to form diverse structures. The method leverages a virtual-force framework where robots attract toward dock-connected neighbors and repel from non-connected ones, requiring minimal onboard sensing and no external infrastructure. Simulations show effective localization during the assembly of towers and cantilevers, enabling accurate, scalable, and free-form self-assembly.