Reservoir has launched an agricultural robotics incubator for Washington growers, incorporating researchers and growers to work on solutions in orchards and vineyards. This initiative addresses rising labor costs and workforce shortages for automating high-value crops like apples and wine grapes. Reservoir functions as an agricultural innovation center and venture capital fund focused on field-based solutions for agtech startups. The organization opened its flagship on-farm innovation hub in Salinas, featuring test fields and innovation barns. Reservoir plans to expand this on-farm innovation center model across other key regions in California and the American West.
NASA plans to build Space Reactor-1 Freedom, the first nuclear-reactor-powered interplanetary spacecraft, for an uncrewed mission to Mars by 2029. This propulsion method uses a nuclear reaction to generate electricity for ion propulsion, offering high efficiency for long-duration space travel. Despite past setbacks, NASA is pursuing this technology as a $2 billion bet, focusing on nuclear electric propulsion. The spacecraft concept involves a reactor with a 20-kilowatt capacity and fins for heat radiation. Development is advancing, with the Idaho National Laboratory working on the reactor, and the technology may also pave the way for nuclear power on the Moon.
Engineers at MIT and EPFL designed a robot called a flapping-wing aerial-aquatic vehicle (FAAV) capable of transitioning between swimming and flying. This small vehicle uses flexible wings, an electric motor, and a steerable tail to achieve this transition. Experiments determined that medium-sized, flexible wings with specific flapping frequencies and a 70-degree tail angle were optimal for the transition. The research aims to create a drone that can access and sample aquatic regions for oceanographic purposes. Future work will focus on improving wing articulation and testing performance in turbulent conditions.
RoboBusiness has been rebranded as the Robotics Summit & Expo West, focusing on the design, deployment, and scaling of commercial robots and businesses. The new event addresses the industry's shift from prototyping to real-world implementation, focusing on challenges like fleet management, go-to-market strategy, and data-driven product decisions. The summit aims to share lessons from industry leaders on moving products from pilot to deployed, scalable businesses. The event program is structured around key themes including deployment at scale, commercial strategy, field data utilization, and investment growth. The rebrand clarifies the portfolio by separating focus between the Boston event, which emphasizes technical challenges, and the West event, which focuses on scaling the company behind the product.
RoboBusiness 2026 celebrates the 20th anniversary of the event by focusing on the past two decades and future trends in robotics. The event features keynotes from industry leaders discussing topics like AI in robotics and production, and various breakout sessions. Attendees can engage with demonstrations and talks on topics ranging from data capture and supply chains to humanoid motion and physical AI. The schedule includes a Women in Robotics Luncheon and a Networking Reception. Numerous sessions are planned in both the main hall and the Engineering Theater throughout the day.
Open Robotics, Saha Robotics, and Terminal Istanbul will co-host the first ROSCon Türkiye 2026. The conference is scheduled to take place in Istanbul, Türkiye, from December 3 to 4, 2026. It will be conducted in both English and Turkish. Submissions for talks can be made via the website's Talks section. Further information and ticket registration are available on the official website.
Voxel51 launched version 3.0 of its data platform to help developers manage large amounts of sensor data from robotics and physical AI systems. The platform integrates tools to search, organize, label, and evaluate this data, addressing the difficulty of manual data management for growing fleets. It features Voxel51 Agent, an AI assistant that automates tasks like finding and labeling relevant data. The system includes multimodal time-series search and agentic labeling capabilities to analyze synchronized sensor streams. The platform aims to transform robot experience into training data by efficiently extracting useful examples for AI model improvement.
Indoor autonomous mobile robots succeed by operating in environments specifically engineered for them, relying on a foundational assumption of controllable spaces. Outdoor autonomy requires robots to adapt to highly variable, unengineered environments, which presents fundamentally different engineering problems. Real-world outdoor deployment is constrained less by perception and more by the system's ability to manage unpredictable transitions and maintain reliability across diverse conditions. Successful outdoor robots must incorporate human coexistence from the start, prioritizing behavioral predictability over mere technical capability. The next decade of AMR development will be defined by platforms designed for unstructured environments, not just those that can operate within engineered settings.
Power supply issues are often the root cause of robotics failures, despite teams focusing on software and perception. Voltage sags and electrical noise from motors and drivers frequently cause performance problems that software alone cannot fix. Understanding power architecture is crucial because it dictates the reliability of sensors and compute modules. Datasheets must be read defensively, focusing on load response, voltage dropout, and thermal derating rather than just efficiency. Reliable robot design requires treating the power stage as a first-class design problem to ensure system stability.
Exotec expanded its warehouse robotics business in Japan by installing two Skypod systems with Mitsubishi Shokuhin and Trusco Nakayama. These deployments, in partnership with IHI Logistics Industrial Systems, enhance automation in the logistics sector. Mitsubishi Shokuhin used Skypod to improve picking efficiency and automate operations at its food distribution center. Trusco Nakayama implemented the system at its logistics center to automate order fulfillment for industrial supplies. The projects demonstrate the growing adoption of warehouse automation in Japan to address challenges like labor shortages and productivity demands.
Eka Robotics, founded by MIT professor Pulkit Agrawal and Tuomas Haarnoja, developed a manipulation architecture combining a force-sensitive gripper and a new deep-learning model. This technology allows the robot to adapt to a large number of objects without retraining, demonstrating superhuman performance in dexterous manipulation tasks. The system uses a vision-force-action model, focusing on physical interaction, which allows the robot to avoid crushing objects by reacting to touch. Eka can pick up objects rapidly, and its method leverages simulation to generate vast amounts of robot experience. The company aims to bridge the generality-performance gap in robotics by integrating physical understanding with high-level intelligence.
Dream Labs has released its Dream-1 robot API, with pricing starting at $3 per hour, according to Mete Gultekin. Gultekin claims the Dream-1 model can control various bimanual robots without requiring specific fine-tuning for each robot. He observed a trend where robotics is entering an "API era," with developers offering access to robotics models. Other models, like Google DeepMind's Gemini Robotics ER 2, use token-based pricing which is not directly comparable to the hourly robot API cost. Gultekin emphasized the importance of reliable, repeatable task performance across different hardware and environments.
DexGraspFlow is a framework that uses topology-aware SE(3) flow matching to generate dexterous grasps across different embodied systems. It integrates geometric relations and kinematic topology to address the challenges of cross-embodiment grasp generation. The model employs dual-stream Diffusion Transformers to model link-link and link-object coordination, incorporating attention biases from kinematic-tree distances. The framework learns velocity fields to generate link poses, which are then inverted to recover joint configurations. Experiments show high success rates and reduced latency compared to state-of-the-art methods, demonstrating zero-shot adaptability to variations.
Embodied VLMs need replanning for error recovery when execution mistakes occur during complex manipulation tasks. The proposed method introduces the Embodied Replanning for Recovery Environment (ERR-Env) to provide verifiable feedback on task progress and errors. Action-Grouped Turn-GSPO (AGT-GSPO) is an online reinforcement learning technique that uses turn-level rewards and error states to train an embodied VLM. This method is applied to train an ERR-VLM on an ERR-Bench benchmark featuring nominal, intermediate, and error states. Evaluation shows that AGT-GSPO significantly improves the VLM's success rate for error-free completion and recovery.
Embodied Playground is a benchmark designed to evaluate foundation-model agents' ability to coordinate physical, strategic, and spatial reasoning over long task horizons. Agents must track progress, adhere to constraints, and adjust plans based on execution feedback during these tasks. The playground encompasses 410 environments covering manipulation, puzzles, and household activities with varying complexity. It includes curated subsets like EP-100 and EP-Easy/Hard for structured evaluation. The framework provides a unified method to study sustained reasoning and action in embodied robotics.
General-purpose multimodal models show strong spatial perception but require stronger grounding for physical tasks. The research develops a unified data pipeline to create verifiable grounding supervision by combining robot recordings with visual data. Fine-tuning Qwen3.5-9B on this supervised data results in ACE-Brain-0.6. This model outperforms GPT-6 Astra on several grounding benchmarks, demonstrating improved spatial understanding. Scaling verifiable grounding supervision effectively strengthens the spatial perception of foundation models for embodied operations.
Reliable contact establishment is a challenge for autonomous orbital servicing using gecko-inspired dry adhesives. This paper introduces a framework to estimate the contact state of these interfaces using embedded capacitive sensing during the adhesion lifecycle. Five interaction states (idle, proximity, contact, engaged, and disengaged) are classified using sensor data and evaluated with six machine learning models. The Random Forest model achieved 95.65% offline accuracy, improving to 93.14% online accuracy when combined with a rule-based state machine. Incorporating physics-consistent state constraints enhances robustness for real-time interaction-aware capture in space missions.
A novel method is presented for enabling semi-autonomous robots to recognize task goals and plan actions for educational tasks. This research focuses on "Shoebox Tasks," which are visually defined activities used for educating individuals with Autism Spectrum Disorder (ASD). A solver utilizing a vision-based Large Language Model converts task images into a Planning Domain Definition Language (PDDL) representation and generates action plans. A robotic interface allows a robot, like a Nao robot, to autonomously execute these tasks and provide assistance through demonstrations or commands. The system aims to enhance robot-assisted learning opportunities and foster interest in robotics for students with ASD.
Diffusion policies with linearly composed scores are used for multitasking robotic learning through separate factor combination. Current methods fail to encourage component specialization, leading to uninterpretable denoisers that average identical predictions. The authors propose a drop-in score-decorrelation auxiliary loss, based on projective composition theory, to encourage specialization. This loss penalizes overlap between component noise predictions, promoting specialization without architectural changes. Empirical results show this loss reduces double-counting and improves task decomposition while maintaining multitask performance.
EmbodyHOI generates hand-object interactions for humanoid dexterous manipulation by incorporating embodiment-aware constraints into the trajectory generation process. It achieves this by running a differentiable inverse-kinematics solver within the sampling loop of a flow-matching trajectory generator. Gradients from this embodiment-aware solver guide the HOI sampler, jointly optimizing the humanoid's position relative to the object. This method significantly reduces the retargeted wrist error and improves task success rates compared to other embodiment bridging techniques. EmbodyHOI demonstrates superior performance in bridging the embodiment gap for robotic manipulation tasks.
XoNav is a four-actuator hip exoskeleton designed to provide personalized haptic guidance for assisted navigation. It delivers advance cues, braking cues, and steering cues directly at the hip joints based on gait-state classification from IMUs. A bilateral LSTM classifier identifies gait states, allowing cues to be delivered during walking while suppressed during standing. Bayesian optimization personalizes braking for minimum resistance and steering for maximum comfort while ensuring safety. Experimental trials showed participants successfully navigated blindfolded paths with improved speed and accuracy.