People become more suspicious of expressive humanoid robots when they make conversational errors, especially when they are supposed to be engaging conversation partners. A study involving 50 participants interacting with the expressive robot Pepper found that errors increased oxytocin levels, which correlated with decreased trust and less reliance on the robot's advice. This effect was linked to brain activity in the dorsolateral prefrontal cortex and the medial prefrontal cortex, indicating how the brain processes social awkwardness. Animated robot errors were perceived as social violations rather than mere technical malfunctions, affecting trust. This research suggests that expressiveness in robots influences how people judge mistakes and ultimately affect their trust in the technology.
Existing geometric reconstruction methods only capture relative geometry, leading to inconsistencies in predicted object dimensions across different views and scales. Metric-Aware Geometry Perception (MAGP) is proposed as an end-to-end framework to reconstruct metric geometry that aligns with real-world scale. The framework uses Metric Scale Equivariant Augmentation and Flexible Metric Conditioning to ensure metric consistency from camera and depth observations. Experiments show MAGP reduces absolute error significantly and improves performance in robotic policies. Integration into robotic policies yields performance gains on various benchmarks like RoboTwin.
A study evaluated remote human-robot interaction via virtual reality for inspecting leaves and assessing soil moisture in a greenhouse. The system used an unmanned ground vehicle and a robotic manipulator for these tasks across fourteen plants. Leaf inspection achieved up to 88% disease detection accuracy, while soil moisture assessment successfully determined watering needs for some plants. The reliability of soil moisture assessment was strongly dependent on plant canopy morphology, with dense canopies posing sensing limitations. This research suggests adapting sensing strategies to canopy density is necessary for improved system robustness.
Scaling robot data is challenging because embodied collection is costly and sparse, making representation quality a bottleneck for Vision-Language-Action (VLA) models. The paper proposes VLAct, a VLA-oriented backbone that uses broad, heterogeneous robot data for continued pre-training. VLAct preserves the Vision-Language-Action (VLA) prior and encourages shared action semantics across different embodiments through specific training techniques. VLAct consistently improves downstream performance across simulation, real-world, and unseen-embodiment transfer. This method demonstrates that representation-centric pre-training offers a valuable axis for VLA progress beyond pure data scaling.
Vision-language-action models often lack explicit reasoning over future task dynamics necessary for fine-grained manipulation. PHR-VLA is a framework that addresses this by enabling planning-horizon reasoning in VLAs using privileged latent representations of future dynamics. It introduces a lightweight auxiliary future head to align internal representations with latent dynamics extracted from future observations during training. Supervision using local, contact-centric patch-level dynamics significantly improved performance on manipulation tasks. These results show that aligning latent dynamics effectively provides a strong training signal for improving anticipatory reasoning in VLA policies.
Autonomous navigation in subarctic forests is challenged by dense tree canopies and atmospheric attenuation, making GNSS/cloud reliance precarious. A year-long field deployment revealed that established exteroceptive modalities struggle with seasonal changes and self-similar scenes. Complex SLAM algorithms offer limited accuracy improvements and increased system fragility compared to proprioceptive baselines. Visual-based SLAM methods were particularly sensitive to seasonal variations, and cross-season localization with lidar succeeded where radar and visual methods failed. The study concludes by detailing lessons learned from a multi-season evaluation of navigation pipelines.
The paper proposes an iterative Linear Quadratic (LQ) game framework for coordinating motion planning in multi-arm robotic systems. Each manipulator acts as an independent agent optimizing its goals while considering shared states and collision constraints. The method solves local LQ games using Riccati backward recursions to find feedback Nash strategies. Differentiable penalties are incorporated to handle self-collision and inter-arm collisions during trajectory generation. Experiments confirm that this framework yields smooth, safe, and efficient trajectories in high-dimensional multi-robot settings.
Autonomous underwater cave navigation is addressed by a framework using a Vision-Language Model (VLM) with Chain-of-Thought reasoning. This system infers navigable directions from multimodal environmental cues, including light gradients and geometric complexity. The input data comprises RGB imagery, depth maps, and sonar vertical-clearance measurements. Simulations confirm the framework successfully completes end-to-end traversals without collisions in various cave topologies. This method aims to overcome limitations of traditional visual or sonar-based navigation in underwater caves.
Manipulation in partially observable environments requires planning despite incomplete scene information. ROBUST TAMP is a modular LLM/VLM-guided framework designed for reactive task and motion planning where objects may become visible during execution. The framework restricts the planner to visible states, validates actions, and routes them through execution adapters. Object discovery triggers replanning events, and the system reconstructs the scene to update planning based on completed actions. Evaluations compare text-only and VLM planners across various manipulation tasks with hidden objects.
The research studies learning omnidirectional seated locomotion for humanoid robots on a passive mobile chair with casters. The system requires unfixed pelvis-seat contact and intermittent foot-floor propulsion to achieve this motion. A policy was learned using proprioception and velocity commands, without explicit contact sensing or chair state information. Experimental results showed that learned policies could track omnidirectional commands and outperform standing policies in velocity tracking. The method demonstrated zero-shot sim-to-real transfer to a Unitree G1 robot, achieving omnidirectional seated locomotion.