The NVIDIA Seattle Robotics Lab and Isaac team developed robotic systems to assemble GB300 tester trays by tackling busbar assembly and multi-connector insertion tasks. They achieved high success rates by employing a modular pipeline using classical engineering methods, imitation learning, and specialized perception frameworks. For multi-connector insertion, a new framework called Deep Object Pose Estimation Revisited (DOPER) was developed to reliably estimate the pose of connectors. This approach proved more effective than end-to-end learning for tasks involving physical deformation and tight clearances. The work also emphasized the importance of robust robotics infrastructure, such as the TALOS system, for accelerating real-world deployment.
UBTECH and FAW-Volkswagen are partnering to develop and test humanoid robot applications specifically for factory logistics. This collaboration builds upon previous work where UBTECH's robots were used for quality inspection at the automaker's Qingdao factory. The joint plan focuses on creating demonstration applications and promoting humanoid robot solutions for smart manufacturing. Currently, the agreement involves joint development and testing, without specifying order sizes or deployment timetables. The partnership aims to explore how these robots can reliably perform logistics jobs within the automotive manufacturing environment.
REK is adding swords to its humanoid robot fighting event, which is scheduled for October 9 in San Francisco. The event will feature blade-equipped machines alongside previous robot confrontations. The event card includes fights such as a G1 weapon fight, a T800 bout, and a Rektile versus T-REK match. REK CEO joked that swords were added because the robot fights were deemed too dangerous. The event will feature arm-mounted weapons like Spartan Blades and Claw Blades.
An open-source ROS 2 bridge for the Rokoko Device SDK has been released, designed to publish solved hand data from Smartgloves II directly into ROS 2. It provides 26-joint hand poses, optional TF transforms, and RViz visualization of the hand skeleton. The bridge also includes per-hand diagnostics and a calibration service to define the forward direction. A single launch command manages the driver, hand solver, and bridge, supporting one left or right glove at a time. The software requires ROS 2 Lyrical Luth and Rokoko Device SDK 0.10.0 or newer.
Graph-based world models, utilizing factor and scene graphs, provide a foundation for internal world representations. Factor graphs offer uncertainty-aware geometric estimations, while scene graphs facilitate structured semantic abstraction. The review addresses how to incorporate dynamic aspects of real-world environments into these models. It assesses representation types, construction/update pipelines, and exploitation strategies, emphasizing hybrid models. Open challenges include uncertainty propagation, dynamic motion observability, and the need for robust evaluation in dynamic settings.
SAFE is a low-cost method for real-time slip and fracture detection in robotic grasping using passive acoustic sensing and motor proprioception. It utilizes two polyvinylidene fluoride (PVDF) acoustic sensors mounted on a compliant gripper to gather data without needing vision or prior material knowledge. A HistGradientBoosting classifier processes a 79-dimensional feature vector to report the grasp state as normal, slip, or fracture. Experiments show the system achieves an Alert-F1 of 0.884 and enables an adaptive grasp controller with high success rates across diverse objects. This approach allows for robust, failure-aware grasp control without requiring object-specific calibration.
A context-aware adaptive spraying framework using Vision-Language Models (VLMs) is proposed for agricultural robots to optimize pesticide application. This framework integrates information on crop type, pesticide type, and environmental data like weather into adaptive spraying decisions. A trajectory-tracking controller based on Model Predictive Path Integral (MPPI) control is used to ensure precise navigation and spraying. Comparative results show the method improves crop row detection accuracy by at least 30%. Furthermore, the system demonstrates the ability to flexibly adjust spraying volume and speed while minimizing pesticide drift.
Go2-DrivoR adapts the DrivoR framework for end-to-end trajectory planning for quadruped robots navigating urban environments. It conditions trajectory generation on a local-frame subgoal using a goal token and modifies the vehicle-centric scoring formulation. The method redefines drivable-area compliance for sidewalk navigation and reformulates the ego progress term to be goal-conditioned. Training on TartanGround simulation data allows Go2-DrivoR to improve waypoint-conditioned planning in new simulations and enable zero-shot real-world trajectory prediction.
A robotic platform was introduced for autonomously navigating liquid droplets on an open surface using model-based reinforcement learning. The system uses a two-axis tilting board and camera for real-time feedback to manage complex droplet dynamics. The learned policy was trained on a small number of physical episodes without prior simulation or analytical models. The platform demonstrated the ability to robustly transfer to unseen geometries and autonomously discover strategies for depinning stuck droplets. These results advance reinforcement learning for manipulating deformable soft-matter systems in self-driving laboratories.
CIRRA is a dual-level framework for reconciling new instructions with ongoing execution for embodied robot agents in household tasks. It combines LLM-based semantic reasoning with rule-constrained structural integration to handle new user commands. The framework grounds instructions to executable skills while preserving the ongoing task sequence as an execution backbone. It uses a semantic reasoner to evaluate modified segments and select the most logically coherent local integrations, mitigating ambiguity. Experimental results show CIRRA significantly outperforms baselines in decision agreement and execution on humanoid robots.
Wearable interfaces enable operators to collect dexterous demonstrations by directly interacting with objects, making the hardware part of the data generation process. Interface geometry influences both task performance and the recorded tactile observations. Experimental results show that revised interfaces and spatial-plus-force input outperform baseline methods for tasks like lid twisting and grasping. This indicates that tactile interactions and interface characteristics should be evaluated alongside operator experience. Therefore, wearable interfaces must be assessed based on the recordable tactile interactions and the resulting policy performance.
Deep-space crews require a method for anomaly triage without real-time ground support, as evidence may be inaccessible. The proposed approach uses an evidence-driven architecture for human-agent-robot teaming in Earth-independent scenarios. This system employs a triage state manager to track hypotheses, evidence, uncertainty, and context, with an embodied agent eliciting observations. A mobile robot acquires localized evidence to distinguish between potential causes, as demonstrated in deep-space and lunar base scenarios. The core contribution is a closed evidence-loop architecture tested in a hardware-in-the-loop prototype.
Learning whole-body forceful loco-manipulation for humanoids requires forceful interaction across the entire body due to shifting center of mass and sustained loads. The HULK framework uses model predictive control (MPC) to guide reinforcement learning with predictions of loaded dynamics. Two teachers are trained: one for arm motions under wrist forces and another for locomotion while holding heavy objects. A capture-point control barrier function is used to improve balance, and these teachers are distilled into a single policy. Evaluation showed the barrier function teacher significantly reduced tracking errors and motion divergence compared to MPC-RL alone.