Generalist AI has introduced GEN-1.5, a robot foundation model that can learn new physical tasks from a single demonstration of 3 to 12 seconds without gradient updates. The model achieved a 59% average success rate (±10%) in 10 diverse tasks and 83% (±9%) after 10 gradient steps on five minutes of data per task. GEN-1.5 exhibits zero-shot sim-to-real transfer, allowing demonstrations recorded entirely in simulation to successfully prompt behaviors on real-world hardware. The model's efficiency is attributed to relentless pretraining and its use of handheld grippers, marking the end of traditional teleoperation as a data-collection method.
Unitree CEO Wang Xingxing admits the company has deliberately held back from large-scale factory rollout due to technical limitations. The machines can perform simple assembly, but their throughput remains lower than a human worker's and requires retraining from scratch for every new task. The main bottleneck is the "lossy" nature of physical robotics, which cannot correct tiny deviations in the final millimeters of a task.
FastSwarmSim is an open-source, lightweight multi-UAV simulator for PX4-compatible vehicles, designed for multi-robot algorithm development and repeatable simulation-time experiments. It combines a trimmed PX4 runtime, MAVROS-compatible ROS interfaces, RViz visualization, and a conservative lock-step simulation clock that advances the simulation in lock-step, making simulation progress explicit and reproducible. The simulator can achieve approximately 100x real-time acceleration, depending on the vehicle count, and is useful for high-volume algorithm testing, repeatable multi-UAV experiments, and timing-sensitive coordination studies.
HM-LD1's 3D depth camera now officially supports ROS1 Noetic and ROS2 Humble/Lyrical, with an open-source repository available. The ROS1 Noetic & ROS2 Humble/Lyrical driver includes support for Serial, UDP, and UVC transports, as well as depth and point cloud data.
Researchers and developers using ROS across multiple machines or LANs use various networking solutions, including OpenVPN, Tailscale, ZeroTier, self-managed WireGuard, and other VPNs, to enable remote access and communication between robots, ROS 1/2, and backend systems.
Waymo's onboard computing system powers its autonomous driving technology with a custom 5 nm ASIC delivering over 1,000 TOPS of machine learning performance. The system, designed to provide millisecond-level response times, operates under vibration, temperature extremes, and other demanding conditions. The architecture combines Nvidia accelerated computing with technology from AMD, Micron, Samsung, Sandisk, Socionext, and TSMC.
Logic Robotics has launched an autonomous pallet for rail-to-road transloading, powered by the Logic Digital Twin Operating System, to streamline the process of transferring goods between railcars and trucks. The system integrates autonomous mobility with real-time digital twin visualization to eliminate inefficiencies caused by manual handling and forklift-based processes. The result is faster throughput, improved safety, and reduced operational downtime across intermodal terminals.
APPROVE is an LLM-based multi-modal end-user programming framework for robot programming, integrating natural language input with a block-based interface and explicit user confirmation. Generated programs are visualized in Blockly, allowing users to confirm, modify, or reject them before execution. Confirmed functions are stored in a library for reuse, building a set of reliable program components. This approach emphasizes user trust, intent alignment, and reusability.
A two-degree-of-freedom bilateral force-feedback telemanipulator paired with a 32-DoF tactile fingertip display is used to study teleoperation tasks. The results show that reproducing distributed contact information significantly accelerates task performance and brings teleoperated movements closer to natural human behavior, reducing the deviation by 27.9%. Increasing the resolution of tactile feedback further compresses the state-space distribution of teleoperated motions, associated with improved training outcomes for autonomous robot policies.
A neural reduced dynamics (NRD) framework is developed to learn the "right abstraction" - a reduced state that preserves control-relevant physics of high-fidelity systems. This framework separates the state the model propagates from what can be supplied as an input or recovered analytically, and trains policies entirely inside the frozen learned model. The framework is validated on two case studies across three control tasks, achieving lower median and mean tracking error than single-terrain specialists on all three terrains, including zero-shot bumpy terrain.
SCAPE (Scenario-Conditioned Simulation-Augmented Policy Evaluation) is a framework that predicts scenario-conditioned real-world policy performance using limited paired sim-and-real samples and large-scale simulation rollouts. SCAPE corrects sim-to-real bias in simulation labels before training the prediction model and calibrates prediction uncertainty through conformal prediction. This approach reduces scenario-level prediction error by 4.9% to 27.7% compared to scene-conditioned neural and aggregate statistical baselines, and improves testing sample efficiency and generalization to out-of-distribution scenarios.
Hybrid Feedback Sampling for Sample-Efficient Model Predictive Control investigates the instability of shooting methods in sampling-based MPC and proposes a hybrid sampling design that balances local and global search based on system stability and computation budget. This approach, referred to as Feedback Sampling MPC (FS-MPC), achieves faster convergence than standard MPPI and better optimality than standard feedback sampling. FS-MPC successfully tackles dynamically unstable tasks, outperforming standard sample-based approaches and feedback policies alone. It is validated on humanoid robot locomotion and manipulation tasks in the real world.
SAM-TD, a compilation method, is presented to enforce arbitrary Linear Temporal Logic (LTL) specifications in stream-based Task and Motion Planning (TAMP). It translates LTL constraints into automata and embeds regressed automaton guards into action schemas, allowing SAM-TD to handle objects generated by streams during planning. Experimental results demonstrate SAM-TD's effectiveness in three robotics PDDLStream environments and show competitive performance with standard discrete PDDL benchmarks.
Fine-tuning vision-language-action (VLA) models with self-demonstrated generative control for multi-task manipulation improves performance on target tasks while preserving original instruction following and behavioral priors. Finetuning on in-domain expert data from a new robot causes severe performance drops. A self-supervised method generates online interaction rollouts from the zero-shot VLA as additional training data for finetuning, yielding strong multi-task policies.
A parameter-free approach to LiDAR odometry, LF-GICP, tackles degeneracy in geometrically degenerate environments. It bypasses traditional approaches by using a voxel-normal localizability field, detecting directional anisotropy, and distinguishing information absence. This results in the lowest KITTI relative translation error (0.865%) and generalizes across four sensor types without re-tuning.
A multimodal trajectory planning framework for autonomous surface vehicles is presented, integrating model predictive control (MPC) with a turning circle-based control barrier function (TC-CBF). The TC-CBF accounts for the nonholonomic motion and finite turning capability of a surface vehicle, generating distinct left- and right-turning avoidance modes. This method achieves higher success rates, fewer safety violations, and smaller residual violations than single-mode baselines across all tested traffic densities.
HiTac-WAM is a hierarchical tactile world action model that forecasts a sequence of future tactile states for each candidate action chunk before execution. The model forecasts contact state, a 3D deformation field, and slip risk, organized as a directed hierarchy. HiTac-WAM achieves a mean contact F1 of 0.921 and improves real-robot success rates in tasks such as chip grasping, blackboard erasing, and USB insertion.