AI agents in 2026 are evolving into durable systems equipped with memory, skills, and physical capabilities. Key developments include local agent platforms like OpenClaw, which establish personal AI assistants as file-backed infrastructure. The landscape is shifting towards advanced models such as VLA models for robotics and Web World Models. Infrastructure components critical for these agents include identity management and tool execution via central gateways. A focus on Responsible AI is also emerging alongside these technical advancements.
Harald Braun, CEO of Guident, outlines recent mandatory changes impacting the autonomous vehicle (AV) industry in California. These mandates specifically concern ticketing systems, operational geofencing requirements, and a minimum mileage threshold of one million miles. The post details how these new rules establish the current compliance landscape for driverless technology deployment.
NVIDIA and Hugging Face have introduced new models and frameworks to LeRobot. LeRobot is an open-source library designed for the robotics field. It facilitates training, running, and sharing essential robotic components. These components include datasets, machine learning models, policies, and entire workflows. The announcement regarding these additions was originally published on The Robot Report.
ENCY Software and Stäubli Robotics have formed a global partnership with the goal of improving robot programming. This collaboration aims to make the process more intuitive, faster, and generally more accessible for users. The agreement signals a joint effort to streamline complex robotic setup tasks. Details regarding this partnership were initially reported by The Robot Report.
Traditional industrial automation involves engineering rigid systems for specific, unchanging tasks, which makes adapting to new processes difficult and costly. Current market demands challenge this fixed, single-purpose approach from manufacturers. Intrinsic proposes a shift towards building a "software-defined factory." This new model aims to overcome the limitations of existing, inflexible automation hardware. The goal is to achieve greater adaptability in manufacturing environments.
Previously, labeling Serrano hams required significant manual labor because identifying bone placement was necessary for tracking aging and drying stages. A Spanish systems integrator has engineered an automated solution to address this difficulty. This newly developed system utilizes artificial intelligence and robotics. The robot can now successfully perform the crucial task of ham labeling. This marks a significant automation achievement in cured ham processing.
Eclipse Automation debuted Eclipse RealitySync at Automate 2026. This new immersive platform aids manufacturers in evaluating and de-risking factory automation projects prior to physical construction. It allows exploration of virtual production environments utilizing Apple Vision Pro technology. Teams comprising engineering, operations, and leadership can use the system to review detailed factory layouts. Ultimately, RealitySync helps streamline the planning and scaling of complex automation initiatives.
Doosan Robotics introduced PalletizHD+, an AI-powered palletizing solution intended for high-mix manufacturing environments. This new system aims to automate the process of creating stable product pallets while minimizing required engineering effort. It was showcased at Automate 2026 and operates on Doosan's proprietary PalletizOS platform. The technology is designed to increase production flexibility, accommodating varying requirements easily. Ultimately, it offers manufacturers a complete, ready-to-use solution for complex palletizing tasks.
The manufacturing sector faces growing pressure to increase output, reduce waste, adapt quickly, and manage costs. Industrial automation is central to navigating these industry transformations. Suppliers supporting this automated evolution are also playing a critical role in modernizing factories. JLCMC is emerging as a notable provider in the industrial automation supply space. The company's services aim to help redefine solutions for today's advanced manufacturing environments.
NativeMEM is a new Vision-Language-Action (VLA) policy designed to maintain long visual histories efficiently for robotic manipulation. It employs Native Memory Compression, an encoding scheme that compresses historical frames from multiple camera views into single tokens using the VLA's existing vision encoder. These memory tokens are appended to the input sequence, allowing the pretrained VLA to attend over extended history with minimal latency overhead. The system requires no external planner or new memory modules, enhancing simplicity and efficiency. Evaluation shows NativeMEM significantly improves success rates across simulations (32.4% to 84.0%) and real robots (up to 98.7%) while using only 20% of the data required by prior methods.
RoboSnap is a novel framework designed to convert single RGB images into simulation-ready scenes for generalizable robot learning. It employs a layered architecture that distinguishes between physics-critical interaction areas and the surrounding visual context. Collision-aware foreground assets ensure stable physical interactions, while 3D Gaussian splatting maintains visually accurate backgrounds from novel viewpoints. Evaluations on DROID scenes demonstrate RoboSnap's ability to enable reliable trajectory replay and support synthetic data generation. Additionally, the authors introduce DROID-Sim, a new dataset comprising 564 real-world scenes to advance research in real-to-sim transfer.
Multi-agent collaborative SLAM relies on exchanging global descriptors for inter-agent loop closure detection, but encrypted radios fail if a swarm member is compromised. A corrupted agent can reconstruct an honest agent's imagery and trajectory from public descriptor broadcasts. To mitigate this privacy threat, CILC proposes using Secure Multi-Party Computation (SMPC) specifically for detecting ILC candidates. This approach allows similarity comparison without exchanging raw global descriptors in the clear. Evaluations show that CILC remains real-time and feasible across visual and LiDAR data while preventing information leakage to malicious agents.
Vision Language Action (VLA) models integrate visual perception, language understanding, and action generation into one framework, enabling robots to follow complex instructions from images. These models are increasingly dominant in learning-based manipulation, with bimanual coordination proving a highly challenging application area. Unmanned aerial robotics presents a parallel challenge requiring coordinated control of thrust, attitude, and grippers from visual input while managing latency. This review surveys 183 contributions across VLA architectures, training methods, and action representations. Notably, the strategies developed for bimanual manipulation are transferable to unmanned aerial systems, suggesting fourteen key research directions for both fields.
Presents EvoPlan, a neuro-symbolic framework for robot planning addressing limitations in both LLM and classical planners. The system first mines global Signal Temporal Logic (STL) constraints from demonstration data to codify rules or preferences. Second, an evolutionary PDDL planner uses LLM proposals, programmatic validation, and iterative plan refinement on benchmarks like ALFWorld. A final constrained execution loop translates the plan into waypoints, checking them against the mined STL constraint and triggering re-planning upon violation. This entire pipeline is designed for local, on-robot deployment using open-weight models.
Modular soft robots offer significant potential for complex applications due to their compliance and reconfigurable structure, but controlling them is difficult because of nonlinear dynamics and morphology changes. Existing controllers require retraining when the robot's physical configuration alters. This work introduces a continual learning framework that allows the controller to adapt incrementally to new MSR configurations without losing prior knowledge. The framework can also learn module-specific dynamics in a distributed manner for fixed setups, improving precision. Validation was performed via simulation and on a three-module pneumatic soft robotic arm, demonstrating adaptive control by selectively activating necessary modules.
G-PROBE is a novel, learning-free framework designed for global localization from 3D point clouds, addressing challenges posed by limited or asymmetric fields of view (FOV). The method utilizes cross-FOV branch ensembles in its front-end to create heading hypotheses for robust place recognition regardless of sensor configuration. A score-scale-invariant component, gamma-SGRT, actively suppresses heading aliasing when the FOV is partial but remains inert at full coverage. The back-end refines pose estimation using a modified GICP that restricts matching only to high-certainty points identified via a bird's-eye-view certainty map. Evaluated across various LiDAR datasets and sensor modalities, G-PROBE outperforms baselines when facing significant cross-sensor pairing or FOV asymmetry changes.
CaLiSym is a novel framework extending exact symplectic learning to practical, non-conservative systems like robotics. Instead of enforcing symplecticity on the measured state, it embeds the system into a structured lifted canonical phase space where the learned dynamics remain exactly symplectic. The lift process is algebraic and explicit, avoiding complex components like latent states or iterative ODE integration. Testing on dissipative pendulums, quadrotors, and quadruped robots showed that CaLiSym consistently improves out-of-distribution prediction using parameter-efficient models. These results validate extending symplectic learning geometry preservation to real-world, energy-exchanging robotic dynamics.