Scientists at Nanyang Technological University, Singapore, have developed a seed-sized surgical robot that can navigate soft and uneven surfaces to perform five surgical functions wirelessly, including cutting biological tissues, releasing drugs, gripping tissue samples, and generating heat remotely. The robot, measuring 4.4 mm in length, uses magnetic coils and soft magnetic materials to control its movements and switch between different functions in under a second. Testing on biological tissues revealed that the robot was able to perform the specified functions with minimal damage or adverse effects.
Dyna-2, a world-action model pre-trained on over one million hours of human video data, demonstrates a human-to-robot transfer scaling law, proving that scaling human video predictably improves zero-shot performance on unseen robot hardware. The model achieved an 87% pass rate in real-world zero-shot deployments, outperforming its Vision-Language-Action predecessor, Dyna-1. Dyna-2's architecture relies on "video co-training" and introduces a novel one-step video generation distillation pipeline to reduce inference latency by two orders of magnitude.
Researchers at Hanyang University ERICA have developed a vertically integrated dual-gated tribotronic transistor architecture that offers tuneable sensitivity, reliable contact and proximity detection, and minimal pixel footprint for high-density array integration. This technology can pave the way for electronic skin systems that allow robots and prosthetics to reliably perceive human touch. The device exhibits stable response and recovery times of 127 and 212 milliseconds during each contact-separation cycle, making it suitable for applications in healthcare robots, health monitoring, and autonomous systems.
PlusAI has reached 93.4 percent Safety Case Readiness, ahead of its planned commercial launch of factory-built driverless trucks in 2027. The company's Autonomous Miles Percentage increased to 99.8 percent, and Remote Assistance Free Trips rose to 85.2 percent during the first half of 2026. PlusAI expects to reach 100 percent Safety Case Readiness by 2027 and is targeting over 90 percent of trips without remote assistance at commercial launch.
Researchers at the Karlsruhe Institute of Technology have designed a system to predictively disassemble broken robots, salvaging valuable parts while protecting them from damage. The system uses a predictive algorithm and robotic manipulators to break down the robot, with updates made based on observations. The goal is to create a circular economy where old devices are repurposed to make new ones, reducing waste.
Maximo, an AI-powered robotic system developed by AES, has now installed over 180,000 solar panels representing more than 100 MW of generating capacity. The robots using Maximo have exceeded 500 installations per day, while crews using the technology can install up to 24 modules per person per shift hour. Nick Hegeman, chief commercial officer for Maximo, attributes customer adoption to the ease of integration and reliable performance, which has taken robotics from an unproven technology to a practical construction tool.
Researchers at CRAN and LARSEN introduce a novel deep reinforcement learning approach for controlling cable-driven parallel robots (CDPRs) that does not depend on the specific robot configuration. The method trains an actuator-level policy, which controls each motor to achieve its target cable length, instead of traditional methods that learn to control the entire robot. This approach offers two main advantages: a single shared policy can be applied to any CDPR configuration, regardless of actuator count, and reliance on inverse kinematics avoids the more challenging forward kinematics problem.
SC$^{2}$-WM proposes a self-correcting world model framework with internal feedback for closed-loop decision making in Vision-and-Language Navigation in Continuous Environments (VLN-CE). The method derives feedback from world-model foresight to refine state-level plans before action execution. Conditional world-aware adaptation is introduced to selectively update the world model at test time when feedback indicates model capacity insufficiency, improving navigation robustness and generalization in experiments on standard VLN-CE benchmarks.
A self-supervised visual manipulation method learns relative pose corrections directly from wrist-mounted RGB images, generating demonstrations around a target pose without requiring explicit camera to robot extrinsic calibration. The proposed pipeline uses ROS 2 and Isaac Sim to collect labeled image-pose pairs, training a convolutional network to regress relative translation and rotation from single frame observations. In simulation and real-world experiments, the system performs end-to-end grasp attempts with success rates of 66.6% and 63.6%, showing potential for practical visual manipulation with limited setup effort.
A lightweight 2B model equipped with high-fidelity visual inputs successfully matches the overall success rates of massive 7B baselines in Vision-Language Navigation for Unmanned Aerial Vehicles (UAV-VLN). However, this minimalist policy exposes a fundamental robustness flaw inherent to pure Behavior Cloning (BC), leading to alarming collision rates. AeroDPO, a zero-cost automated Direct Preference Optimization pipeline, is proposed to overcome this vulnerability by leveraging deterministic physical simulation state rollback and offline vision language inspection. This approach boosts success rates to 49.16% on unmapped scenarios while drastically suppressing collision rates.
The survey "Learning Physical Interaction: A Survey of Tactile- and Force-aware Robot Learning" proposes a unified hierarchical structure, TF-ART, to map individual methods into a single framework that captures multimodal sensing and multi-phase system design in force- and tactile-aware robot learning. The study characterizes how recent works organize observation modalities, encode and fuse heterogeneous sensory inputs, generate and refine actions across multiple phases, and connect learned policies to reactive robot-end control.
MPPI control handles constraints only through cost, which can lead to inaccurate results. Projection-Retraction MPPI (PR-MPPI) proposes an exact constraint-manifold control for manipulators by projecting sampled velocity to satisfy both equality and inequality constraints. This projection ensures that the returned command satisfies the closed-chain equality to numerical tolerance. PR-MPPI has been validated on 14-DoF dual-arm systems, including a joint-limit stress test and randomized obstacle avoidance.
Researchers have developed self-aligning compound nested lattice modules composed of conjoined cuboctahedral-octahedral units for robotic construction. The modules enable interlocking aggregation along the x, y, and z axes and can be disassembled and reused across different configurations. Compression testing showed a stiffness of 4,556 N/mm, a maximum load of 3,445 N, and a compressive modulus of 17.5 MPa. The system was demonstrated through furniture and architectural scale structures assembled using robotic arms and mobile assemblers.
A unified visuo-tactile-fusion grasping framework, AdaDexGrasp, integrates grasp generation, feasibility prediction, and adaptive refinement to achieve stable and adaptive grasps. The method introduces an efficient visuo-tactile representation that associates tactile signals with finger identities, supporting contact-aware grasp pose generation during planning and tactile-guided refinement after contact. This approach enhances grasp success rates and generalization across diverse objects in simulation and real-world environments.