An AI-driven tool, ROS2-DDSConfig-Optimizer, automatically tunes DDS configurations for ROS2 applications. Previously limited to Fast DDS, the optimizer now also supports Cyclone DDS. The update provides users with an enhanced capability for optimizing their ROS2 communication stack. Demonstrations compare manual tuning methods against the efficiency of using this new tool. Users are encouraged to submit pull requests and report issues related to the tool.
ros2_info is a fastfetch-style ROS 2 workstation dashboard designed as a full-screen TUI written in Rust for headless operation. It consolidates vital debugging information—including live nodes, topics, and build status—into a single view to reduce context switching from multiple terminals. The tool features an integrated PTY terminal and multi-tab editing with Neovim bindings. Additionally, it includes optional offline local AI assistance via Ollama for tasks like error explanation and autonomous code iteration, all implemented with diff-gating. It supports various ROS 2 distributions and offers advanced features such as a graph canvas and sandboxed experimentation mode.
Advanced robotics requires sophisticated management of onboard power sources. Murata Power Solutions' John Quinlan discusses this crucial aspect for humanoid development. His insights cover how power is effectively converted, distributed, and controlled from the battery to the robotic actuators. This technical overview was originally published on The Robot Report platform. Managing this power flow is key to advancing humanoid robot capabilities.
NVIDIA has released insights detailing methods for evaluating general-purpose robot policies for practical, real-world application. This information was initially published on The Robot Report. RoboLab research underpinned the creation of NVIDIA Isaac Lab-Arena. This framework enables large-scale policy setup and comprehensive evaluation in simulation. It provides a platform crucial for validating advanced robotic behaviors before physical deployment.
Gulls, puffins, and petrels are examples of birds capable of both swimming and flying. These species dive underwater to hunt for food before taking to the air. Engineers have been inspired by the abilities of these natural "aquatic aviators." The inspiration stems from how diving birds transition between water locomotion and flight. This bio-mimicry is leading to new developments in robotic technology.
Current agricultural robotics vision systems mainly function during the day, hindering 24-hour monitoring capabilities like nocturnal pest detection. To overcome the lack of nighttime datasets, researchers propose an unsupervised image translation framework converting daytime RGB images to near-infrared (NIR) nighttime versions without needing explicit supervision. This method utilizes a pre-trained CLIP model to maintain semantic consistency across the day-to-night conversion and includes a visibility mask for NIR limitations. Evaluation on the novel AgriNight dataset shows improved performance in downstream semantic segmentation tasks compared to existing baselines. The work supports real-time autonomous navigation testing with physical robots operating at night.
Achieving robust in-hand manipulation without external sensing is difficult due to contact uncertainties and gravity disturbances. The research introduces two physics priors—a global grasp-quality prior and a local contact-geometry prior—to enhance learned rolling stability. The grasp-quality prior functions as a dense reward term, promoting better distributed contacts and wrench resistance. Meanwhile, the contact-geometry prior mechanically shapes the fingertip interface for task-aligned rolling and drift reduction. Evaluations on multifingered hands manipulating various objects show these priors significantly boost rotation efficiency, stability, and disturbance rejection during in-hand rolling tasks.
GaitSpan is a novel framework designed to expand a basic pretrained walking policy to enable faster gaits like jogging and running for humanoids. It treats initial walking as a reusable seed skill that governs essential motor structures like balance and coordination. The expansion involves three mechanisms: rhythm generation using internal clocks, stride shaping based on inverted pendulum dynamics, and residual adaptation for fine motion details. This approach yields a single command-conditioned policy capable of spanning continuous speed ranges across different gaits. Furthermore, GaitSpan demonstrates zero-shot transferability to unseen terrains in simulation and the real world, outperforming multi-expert or imitation learning baselines.
A corpus-level analysis examined behavior diagrams from Sony's R-CODE sample distribution for the ERS-111 AIBO. Instead of analyzing individual scripts, the study compared named states across the entire corpus. This revealed that diverse routines are built upon a compact embodied grammar. The core vocabulary revolves around initialization, sensing, iterative action, synchronization, and recovery. Consequently, this state-based abstraction is proposed as valuable for creating new, modular behavior routines on constrained robotic systems.
Attitude control of spacecraft traditionally relies on thrusters or momentum exchange devices. This research explores utilizing robotic manipulator movements for propellant-free reorientation instead of treating them solely as disturbances. A trajectory optimization approach, incorporating joint and collision constraints, was developed and solved using an interior point solver. Demonstrations show that complex slew and detumble maneuvers are feasible across diverse spacecraft-manipulator systems. The study proves manipulators can function as primary or redundant attitude control actuators, benefiting in applications like in-space assembly.
Robot gaze is crucial for coordinating human-robot dialogue, especially in task-focused settings beyond simple conversation. This study examined how a robot's gaze impacts human attention during a collaborative word association game with an LLM-driven NAO robot. Under conditions of mutual or referential robot gazing, participants' time to first fixation on proposed words was unaffected by the robot's gaze orientation. However, participants were observed directing their gaze more frequently toward the robot when confirmation requests were present in the dialogue. The findings suggest that the cognitively demanding nature of the task may overshadow the specific effects of the robot's referential gaze cues during collaborative work.
DiffRadar is a new, real-time Simultaneous Localization and Mapping (SLAM) system designed for radar sensing. Instead of using discrete scan matching on heatmaps, it models radar observations as a differentiable, physics-aware Gaussian field. By representing the scene with anisotropic Gaussian primitives, DiffRadar uses a differentiable radar forward model to optimize robot pose and scene structure directly from measurements in range-azimuth and Doppler-azimuth spaces. Tested on commodity FMCW hardware, it performs well on benchmarks like Radarize and custom stress tests. The system demonstrates significant reductions in trajectory error and improves map consistency while maintaining 70 FPS real-time performance.
Turbulent aerodynamic interactions severely challenge quadrotors flying in close formations, risking collisions if unmodeled. A proposed physics-informed residual dynamics learning framework models these complex interactions while maintaining the system's differential flatness property. This preserved flatness enables a computationally efficient feedback linearization controller that uses feedforward compensation for disturbances. Hardware tests show this method reduces average tracking errors by 31% versus baseline methods. Furthermore, it matches state-of-the-art NMPC performance using significantly less computation and requires minimal training data.