The Open Source Safety Consortium was started to share safety evidence, not just code, to reduce the need for redoing basic safety analyses. The consortium develops reusable evidence for open-source components, including hazard analyses, safety requirements, tests, and known gaps, versioned with the code. An initial example is a published prototype including firmware, hardware designs, and a draft safety case. While not yet certified, the goal is to provide evidence targeting the rigor of a SIL 3/PLr=e capable solution. The consortium seeks to test how much evidence can be reused and whether the saved effort justifies adapting and maintaining the shared evidence.
Driverless cars are presented as a trend driven by novelty rather than genuine need. The author argues that relinquishing driving responsibilities leads to a loss of essential human skills and purpose. While acknowledging improvements in technology like navigation, he questions the value of these advancements over the skill and cooperation involved in driving. The author suggests that the focus should be on using advanced technology for other significant goals, like energy solutions. Ultimately, he warns that driverless cars risk "lobotomizing" the human experience.
GPT-6 Astra autonomously painted the Golden Gate Bridge by learning to control a robotic arm based on real-time visual feedback. This demonstration sparked a debate about whether the "OpenAI of robotics" resides in general models or specialized physical AI neolabs. Some argue that future value in robotics lies in physical hardware capabilities rather than purely software intelligence. Skeptics suggest that rapid model degradation and the limits of physical contact mechanics will favor hardware-based moats. Ultimately, the discussion centers on whether general models will commoditize physical intelligence or if physical constraints will limit their advancement.
The ROS Meetup Cartagena 2026 is scheduled for November 7, 2026, in Cartagena, Colombia. The event is organized by IEEE RAS Colombia, IEEE RAS Colcaribe, IEEE RAS Javeriana, and UTBot. It aims to bring together students, researchers, developers, and professionals to share knowledge about robotics and ROS. The meeting will take place at the Napoleón de la Rosa Echávez Auditorium at the Universidad Tecnológica de Bolívar. The goal is to connect, learn, and strengthen the open-source robotics community in Colombia and Latin America.
Multimodal demonstrations in robot world models mix operator habit, shared physics, and observation nuisance. Entangled next-observation predictors absorb these three factors. The work formalizes this with a causal model and tests interventions to isolate these components. A habit-aware reverse scoring rule improves low-shot transfer and retains cleaner dynamics. This adaptation rule extends to multi-view and multi-step observations, improving robot performance.
Legged robots are being extended to underwater environments to achieve amphibious mobility for tasks like inspection and monitoring. The research presents the design, modeling, and experimental validation of a reproducible underwater quadruped robot. The robot utilizes custom-machined polyoxymethylene for motor housings and models drag forces to understand torque transmission. A closed-loop attitude controller, based on the special orthogonal group SO(3), was developed to regulate the robot's orientation. The system was successfully tested in simulation and experimentally in a water tank to track desired orientation setpoints.
Vision-guided reinforcement learning for UAVs is investigated using compact target-centric cues and low-dimensional sensor measurements for long-horizon visual servoing. Lightweight target segmentation provides image-space offsets and relative depth, which are combined with other measurements into a compact 12D policy observation. Different curriculum strategies for training, including visual, dynamics, and joint, yielded comparable nominal performance. Proprioceptive measurements and image-space cues proved critical for stable flight and target alignment, while explicit depth was not essential for strong performance. Learned policies demonstrated greater robustness to visual perturbations compared to classical controllers.
Verifiers evaluate robot policies by scoring candidate behaviors, used for both evaluation and training. These verifiers include success detectors, reward models, safety filters, and temporal-logic specifications, surveyed in the paper. Two key properties examined are availability, which relates to the cost and timing of a verdict, and credibility, which measures the reliability of a high score. A finding is that credibility decreases as availability increases, regardless of the source of the judgment. Ultimately, the survey proposes metrics to make verifier claims checkable.
AccelMPC presents an FPGA-accelerated Model Predictive Control (MPC) method designed for high-rate, low-power operation in tiny drones. It achieves this through an end-to-end co-design optimizing the solver, numerical representation, hardware mapping, and physical integration. The method utilizes an FPGA-accelerated Alternating Direction Method of Multipliers (ADMM)-based MPC solver paired with a custom 6g PCB. Experiments show the system can perform 1 kHz constrained MPC with dynamic obstacles, achieving up to 15.6x faster solve times and a 195.4x improvement in energy-delay product. The design, including PCB files, firmware, and FPGA code, is released open source.
Bayesian optimization has emerged over the last decade as a powerful method for automatically tuning controllers and learning in robotics. The paper reviews the state-of-the-art, offering a tutorial on setting up BO and positioning it within broader learning paradigms. It discusses various BO methods and their relevance to control systems and robotics applications. A key future direction highlighted is the need for standardized benchmark problems specifically for control engineering and robotics. The authors propose creating a lightweight benchmark suite and best practices to facilitate further research in this area.
Reinforcement learning often struggles to converge for legged robot tasks where exploration terminates before providing useful gradient signals. The Actuator Dynamics Curriculum is introduced to address this by initializing joint stiffness high and gradually annealing it toward the system-identified value as episodes progress. Using a cart-pole system, the study shows that higher closed-loop joint natural frequency improves the feasibility kernel of the underlying Markov Decision Process. This curriculum was applied to a quadrupedal-to-handstand transition on Boston Dynamics Spot, successfully training a policy for the transition. The work suggests that simulated actuator dynamics can guide curricula for tasks where exploration is limited by termination conditions.
Accurate contact prediction in robotic manipulation requires effective decision-making, which is investigated using a compact, randomly initialized visuotactile world model. Adding touch reduced endpoint-force prediction error significantly across MuJoCo Lift episodes. Tactile persistence yielded lower errors, and matched reward revision increased successful lifting to 93.3% in some cases. The study distinguishes improvements from sensing and reward alignment versus force-constrained control. Findings are based on simulator execution and public sensing records, with limited demonstrated transferability.