Engineers at EPFL and NYU developed ScaFi, a robot modeled after fish, to maintain swimming ability across different sizes. The design uses a flexible tail with proportional fiberglass rods that scale in diameter to maintain the fish-like bending motion. Testing with robots of different lengths demonstrated that the fish-like swimming gait was preserved across scales. While swimming motion scaled well, energy efficiency proved more challenging to scale proportionally. This approach offers a scalable platform for studying swimming performance across various environmental niches.
TinyCVIO is a visual-inertial odometry system designed for nanodrones operating under strict sensing and computational limitations. It utilizes lightweight LED constellations to provide known geometry without prior surveying data. The system tracks LED observations from a millimeter-scale camera at 29.2 FPS, incorporating a rigid-board measurement model for state estimation. Experiments demonstrated that the rigid-board model reduced mean absolute trajectory error by 27% compared to planar points. The system achieved low mean absolute trajectory error and relative pose error during flight tests on a Crazyflie drone.
POIL is a point-based one-shot imitation learning framework utilizing stable dynamical systems for robust manipulation. It addresses the challenge of executing demonstrated trajectories on novel objects under varying conditions by sharing a representation of 3D points on the object's functional part. Point correspondences enable one-shot trajectory transfer, which is grounded by a multi-modal large language model for transfer across viewpoint, pose, and category changes. During execution, multi-view tracking allows a Point-set BCSDM to drive the points in a closed loop by projecting velocities onto a single rigid-body twist. This method extends stable dynamical models to a point set, achieving successful transfer and disturbance recovery in experiments.
The WALT method learns world-model-aligned latent trajectories for autonomous driving by addressing the mismatch between visual world states and geometric trajectories. It achieves this by transferring semantic knowledge from a frozen driving world model into a compact trajectory space using a dual-branch trajectory autoencoder. This approach maps raw waypoints into representations that capture scene-level cues relevant for motion planning. Evaluation on NAVSIM benchmarks shows WALT improves trajectory planning metrics while reducing planner computational load. Overall, WALT demonstrates an effective way to interface world-model-based trajectory planning with action-relevant information.
VkVIO is a method for Visual-Inertial Odometry that uses Vulkan for cross-platform GPU acceleration. It addresses the limitation of previous methods that were restricted to CUDA, thereby improving deployment options. The approach provides state-of-the-art accuracy with causal estimates necessary for real-time operation. The method was successfully demonstrated on diverse hardware, including workstations and single-board computers, outperforming CUDA-based systems. This enables low-latency, low-power, and low-cost VIO applications in robotics and XR.
High-level robotic supervisors rely on reported outcomes to determine subsequent robot actions, necessitating more than just proving generalized reactivity. Designers must address encoding failure-prone capabilities, select appropriate liveness assumptions, and implement auditing strategies before deploying supervisors. The authors present a pipeline for ROS 2 supervisors that generates specifications, analyzes assumptions, audits strategies, and produces executable state machines. Experimental results suggest enumerated encoding with System-Goal and auditing is recommended over other methods for deployability. The auditor is sound for several structural defect classes but is not a general liveness verifier.
DGT-Map develops a self-supervised framework to create directional global traversability costmaps from RGB-D observations and locomotion signals. It uses a shared multi-task backbone to learn common terrain features while employing separate prediction heads for vehicle-specific responses. The resulting costmap is heading-indexed, allowing direction-aware path planning. Evaluation in simulation demonstrated DGT-Map's superior navigation success compared to other baselines on complex terrains. This method effectively addresses the limitations of single, isotropic maps and vehicle-independent estimators.
Policies trained via imitation learning can accumulate errors, causing the robot to drift outside the training distribution. Policy-Calibrated DAgger estimates this offline noise by using the predicted action distribution of recent generative policies. The method measures the spread of predicted actions and closed-loop error relative to expert trajectories. It employs partial denoising and diffusion model properties to unnormalize the error measurement in multimodal action spaces. Experiments show this method outperforms dataset aggregation and matches the best noise level in hindsight.
An LSTM-based pressure predictive control (PPC) was proposed for motion control of a vertebraic soft robotic tail, addressing the complexity of soft robot kinematics. The PPC incorporates inverse kinematics, forward kinematics, and a pressure compensation model to achieve non-static and quasi-static motion control. Simulation results showed a 69.8% reduction in RMSE compared to an inverse kinematics-only model. Training the PPC with prediction data allowed for next-moment action prediction and reduced computation time by 60.9%. This method effectively models the soft tail for environmental interaction.
A memory-aware multi-sensor navigation framework is proposed to achieve efficient and safe autonomous navigation in dynamic environments. The framework integrates LiDAR and RGB perception with online distance-field representation learning. It persistently represents static infrastructure while tracking dynamic obstacles. This allows a stage-adaptive Modulated Control Barrier Function Quadratic Program to exploit observed geometry for obstacle avoidance. Experiments show improved navigation performance and collision avoidance in complex settings.