Lars Janssen expressed enthusiasm for driverless cars as he has an eye condition that prevents him from driving. He views these vehicles as a boon for him and his disabled daughter, who relies on others for support. Richard Connell suggested driverless cars could improve access to local amenities like pubs by eliminating the need to drive or take a taxi. Michael Fuller also supported the concept, suggesting driverless cars could be beneficial for accessing local venues. The piece contrasts the technological advancement with concerns about the motivations behind its development.
MIT scientists developed a reconfigurable, robotic laser laboratory to automate the setup and running of complex optics experiments. The system uses a robotic arm to autonomously assemble optical components and fine-tune their alignment with micron-scale precision. This automation allows for the rapid building, running, and dismantling of experiments, enabling scientists to test materials for applications like solar cells. The system incorporates motorized tools and software to precisely manipulate components and autonomously correct for physical disturbances. This technology promises to significantly accelerate scientific research in fields dependent on precision optics.
Users in the medical community using ROS are invited to provide status updates ahead of ROSCon. One update shared is from Vexev regarding their 8-DOF ultrasound imaging robot. Vexev reported successful clinical trial results for their VxWave System and has enhanced hardware safety. The group is focusing on cybersecurity and FDA trials for their work. The community is also seeking technologists for open roles and interested members for the community.
A framework is proposed to transfer tacit human subjective grasping criteria to robots for handling deformable objects. This addresses the semantic gap between qualitative human expectations and quantitative robotic measurements. The system integrates a Vision-Language Model for supervision with a lightweight predictor to adapt to unannotated grasping data. The prediction model allows for rapid online grasping state estimation from tactile and force measurements. Experiments confirm the framework's feasibility in adjusting grasping force based on human-perceived appropriateness.
Service robots can improve task success and user experience by incorporating real-time feedback mechanisms. Users can provide feedback to the robot by pressing a button when the robot performs incorrectly. Robots utilize this simple interaction to consider their failures and replan appropriate actions to complete the task. Experiments showed that this feedback function did not negatively impact the user experience. Specifically, the feedback improved user satisfaction regarding unexpected behavior and anxiety about the robot's actions.
Overconstrained spatial linkages, despite being compact, are difficult to model using rigid body formulations due to redundant constraints. A flexible multibody modeling framework based on the floating frame of reference is presented to address this issue. The study validates this framework experimentally with 3D-printed prototypes of mechanisms with intentional misalignment. Kinematic parameter identification methods are also provided for mechanisms with ill-conditioned Jacobians. Results show that these overconstrained mechanisms exhibit a self-assembling tendency driven by structural compliance.
The LEAP method learns emergent active perception for quadruped navigation without explicit task objective augmentation. It addresses goal-oriented navigation over hazardous terrains where goals must be discovered visually. The approach uses a gaze-invariant representation integrating depth images into egocentric belief maps to facilitate active perception. Training on a terrain curriculum allows task pressure alone to induce the emergence of gaze control. Validation showed LEAP achieves a 92.7% success rate in evaluation scenarios, outperforming scripted and passive perception methods.
Real-world robotic disassembly demands long-horizon execution involving sequential manipulation tasks with diverse configurations. The proposed framework incorporates task context from language to guide skill selection and link instructions to visual targets in the scene. It combines hierarchical task selection with task-context-aware imitation learning to ground language in visual representations for disassembly. This method achieves generalization across various connector geometries and configurations without needing explicit object annotations. The approach results in significant improvements in end-to-end task success compared to existing baselines.
Vision-Language-Action (VLA) models for autonomous driving use future scene generation for supervision, which incurs training burden and inference latency. RAF-VLA proposes aligning planning-relevant internal representations directly with future-frame representations to mitigate these issues. This is achieved through Future-Aligned Supervised Fine-Tuning, a simple regularization method. Experiments on the NAVSIM benchmark show RAF-VLA achieves competitive planning performance with fewer training samples. Furthermore, RAF-VLA introduces minimal overhead, with only 3.8% training and 1 ms inference overhead.
CALOS is a runtime safety layer designed to enforce attitude constraints on quadrotors in Deep Reinforcement Learning without altering the core learning algorithm. It achieves this by formulating four tilt-angle inequalities and a Lyapunov descent condition into a single quadratic program. This program determines the minimum-norm correction to the policy's nominal torque output to satisfy the constraints. The solution is found efficiently using active-set enumeration over the three-dimensional torque space, allowing for real-time constraint enforcement in parallel environments. Evaluation showed CALOS reduced lateral tracking error by 55-60% compared to unconstrained policies while ensuring zero attitude-constraint violations during training.
HINT-Plan is a novel approach to human intention-aware robot task planning by integrating human intention prediction into robot decision-making. It utilizes Vision Language Models (VLMs) to anticipate human intentions from images and convert them into goal states for task planning. The method uses hierarchical Scene Graphs to represent the environment and translates knowledge into formal planning language for executable plans. Evaluation in simulation showed HINT-Plan achieved a 69.71% success rate, significantly outperforming baselines by up to 35.29%. This demonstrates the effectiveness of incorporating inferred human intentions into multi-agent task planning.