Robot safety assurance must now account for attacks that manipulate a machine's perception, decisions, and actions. Research shows that manipulating what a robot sees, hears, or interprets can redirect its behavior without direct control, affecting the entire system. Attacks can occur across three layers: corrupting intelligence at the source, exploiting system vulnerabilities, and manipulating perception and reasoning at runtime. This highlights the need to integrate cybersecurity into functional safety to ensure safety during the robot's entire lifecycle. Therefore, safety assurance must extend beyond functional failure to encompass deliberate manipulation to maintain safe boundaries.
Dr. Scott Walter proposed the "No Bot of Theseus" challenge to test generalized physical capability across ten track-and-field events using a single, unmodified humanoid robot. This challenge prohibits swapping any hardware components, enforcing strict anthropomorphic constraints like five-fingered hands and human-proportional sizing. The rules also require the robot to operate autonomously, log all performance data, and prevent the use of hidden control instructions. To ensure fair comparison, the robot must also perform self-dressing, and all autonomous actions are subject to scrutiny regarding "Performance Enhancing Instructions." Walter argues this unified test is necessary because current humanoids lack the integrated dexterity and strength for multi-event athletic performance.
Lumos Robotics launched the MOS 2, a wheeled dual-arm mobile manipulator designed as an "AI Worker" for high-mix factory workflows. This system has a combined payload capacity of 50 kg and a compact 0.5-square-meter footprint suitable for narrow industrial aisles. Initial trials at Mitsubishi Electric's facility confirmed the system's capability in tasks like visual inspection and material transport. The robot's success is supported by strategic backing from Mitsubishi Electric Smart Manufacturing Technology. The rollout reflects an industry shift prioritizing the operational stability and economic viability of wheeled platforms over bipedal systems.
Three unexpected issues during development significantly altered the design of a sensor board communicating with ROS 2. The initial plan to run ROS directly on the chip proved unfeasible due to flash constraints, leading to the system now using Cyphal with a host-side ROS bridge. A separate issue involved a chip failing to fully power off, which required a specific reset write to resolve state retention. Furthermore, the importance of accurate timestamps was revealed, necessitating the implementation of per-sample timestamps with sequence numbers to ensure correct data fusion. These findings highlight critical considerations for building reliable sensor communication systems.
The study investigates the inverse kinematics of generic 3R positional robots using Conformal Geometric Algebra (CGA). The approach leverages CGA to gain further insights into the kinematic properties of the Inverse Kinematic Model (IKM). The method directly results in a univariate polynomial for $\theta_2$ without eliminating $\theta_1$ and $\theta_3$. This is achieved by reframing the problem as the intersection of two circles within the CGA framework. Ultimately, the research provides an alternative method for solving the inverse kinematics problem for 3R robots.
The research employs geometric analysis, combined with inverse kinematic models from conformal geometric algebra, to investigate generic 3R robots. The study develops a framework to determine the necessary and sufficient conditions for a 3R robot to possess four inverse kinematic solutions (IKS). This framework is used to examine the distribution of IKS within the robot's workspace. Necessary and sufficient conditions are specifically presented for a class of orthogonal robots.
A method is presented for extracting driving scenarios from egocentric datasets using scene graphs and Linear Temporal Logic (LTL). This involves processing sensor data and HD maps to generate a sequence of scene graphs representing a driving scenario. LTL is then used to formally specify driving scenarios of interest, and an off-the-shelf model checker extracts instances of these scenarios from the data. The approach is evaluated on Argoverse 2 datasets and demonstrates effectiveness against a rule-based benchmark. This technique can be applied to both simulated and real-world driving data.
Autonomous highway overtaking requires decision-making that accounts for complex traffic dynamics and temporal risk. A World Model-based Mixture-of-Experts (WM-RMoE) framework is proposed to address limitations in standard reinforcement learning by incorporating explicit dynamics modeling. This framework uses a learned latent dynamics model for multi-step rollouts and a hierarchical gating mechanism to coordinate safety modules. Integration of a Gaussian Mixture Model helps preserve multimodal maneuvering branches, improving robustness. Experiments show the WM-RMoE framework achieves superior safety compliance, decision stability, and generalization compared to baselines.
Gaze-based manipulation for object selection does not guarantee accurate spatial placement required for arbitrary robot positioning. The research addresses the need for cross-device alignment by developing a markerless interaction framework and a dedicated dataset. A proposed method, Graph-based Reference Selection, is introduced to handle sparse robot references. The study introduces Gaze--Surface Intersection Error (GSIE) to directly measure spatial error, showing that conventional pose metrics are insufficient for task accuracy. Ultimately, the work emphasizes evaluating gaze-based manipulation based on task-level performance rather than just pose accuracy.
The research explores using nonlinear body oscillations to achieve natural quadruped gaits by leveraging mechanical resonance. The study presents a framework to make robot nonlinear dynamics predictable, allowing multi-gait locomotion to emerge from these nonlinear resonances influenced by gravity and elasticity. Using the compliant robot eBert, six nonlinear normal modes were identified in simulation and hardware. Black-box optimization demonstrated that each mode naturally develops into a distinct gait with corresponding speeds transferable to the hardware. This work suggests a foundation for designing more agile and efficient robots by exploiting embodied intelligence.
Language-driven grasp models fail when instructions specify both where to grasp and where to avoid grasping. This work introduces an inference-time framework for negation-constrained dexterous grasping that operates without requiring negation-specific training examples. The method uses Sequential Monte Carlo and classifier-free guidance to steer particle sampling toward the grasp location while pruning paths leading to forbidden regions. A frozen 3D part-grounding model localizes the forbidden area from the language instruction. Evaluation on the NegGrasp benchmark shows the method reduces the violation rate of the strongest baseline from 57.9% to 17.2%.
A system of two wearable pneumatic haptic devices enables continuous, closed-loop, bidirectional tactile interaction at relevant force and temporal scales. Each device channel can function as a sensor, actuator, or both, capable of generating skin indentation or measuring touch. This distributed architecture supports rapid response and wireless pairing for real-time transmission of tactile pressure signals. Human studies showed that this feedback improved teleoperated manipulation by reducing applied pressures and mental workload. Ultimately, the device provides a practical foundation for bidirectional touch interaction in teleoperation by unifying sensing and haptic feedback.
DSG proposes a dynamic 3D scene graph construction framework for embodied agents in changing indoor environments. It achieves this by detecting object changes using a dual-view rendering method on semantic-aware 3D Gaussian scene representations. The framework also includes a spatial relationship reasoning method that leverages multi-granularity visual context for improved relationship identification by large language models. A benchmark called DynTHOR is introduced to evaluate scene graph construction in dynamic settings. Experiments show DSG outperforms existing methods in both object node construction and spatial relationship reasoning.
AM-Bench is a modular simulation suite and benchmark designed for learning aerial manipulation policies. It addresses the unique challenges of aerial manipulation, such as environmental disturbances and coupled dynamics, which are often ignored in ground-supported benchmarks. The suite features diverse robot embodiments, various tasks, configurable disturbances, and standard control systems for comprehensive evaluation. It allows for system-level assessment of the interaction between robot embodiment, control, disturbances, and policy choices. Simulation studies and hardware testing demonstrate the diagnostic value of AM-Bench.