Research from the University of Glasgow explored how visual and vocal designs affect user emotional appeal in social robots. Researchers used augmented reality to rapidly test various combinations of animal-like and anime-style faces and sounds on a robot. The study found that combinations of animal-like faces and sounds were most engaging for users. Findings suggest that aesthetic design choices can significantly influence how users perceive a robot's emotional expressiveness. This research aims to guide the future design of more emotionally engaging social robots.
Flow-based Vision-Language-Action (VLA) policies generate action chunks using a task-agnostic Gaussian source, which lacks local motion continuity. The proposed ProAct framework introduces a world-calibrated approach to make the generative source predictable. It uses a Proposal Expert to initialize generation near demonstrated actions and a World Expert to predict future task-consistent latent states. This process calibrates an anisotropic source to refine coupled motion directions. ProAct improves performance over $\pi_{0.5}$ while significantly reducing denoising steps, latency, and increasing throughput.
SoTa is a low-cost capacitive tactile skin designed to provide full-hand coverage for both humans and robots, sharing a common layout of 202 taxels across finger and palm regions. It utilizes a multilayer design with fabric electrodes for in-house fabrication under $10 per skin. The skin maintains over 97% of its response span after extensive loading-unloading and folding cycles. This shared layout facilitates human-robot co-training without requiring learned cross-sensor mapping. Tactile observations improved in-distribution success over vision-only policies across contact-rich manipulation tasks.
NEEDLEWORK introduces an offline dataset-augmentation algorithm to improve robot learning from demonstrations. It achieves this by adding short, verified action bridges between recorded observations in high-dimensional robot data. NEEDLE identifies connections that bypass suboptimal detours and augment the dataset with failed trajectories using only observational data. A sampling technique is presented to incorporate these bridges into policy training without synthesizing intermediate images. Applying NEEDLE to real-robot tasks resulted in an average success rate improvement of 21 percentage points over the strongest baseline.
Safe whole-body motion is crucial for deploying humanoid robots in unstructured settings, often requiring tracking policies to adhere to runtime safety filters like control barrier functions. Treating the tracking policy and safety filter independently causes fundamental mismatches because filtering alters both the executed actions and the state distribution. The study introduces CoFiT, a filter-aware fine-tuning method designed to mitigate these mismatches in pretrained trackers. CoFiT significantly reduces violation time compared to filter-only training across various constraint scenes. Results show that CoFiT yields substantial reductions in violation time and requires fewer safety filter corrections on hardware like Unitree G1.
SocialVLA is a social perception gateway designed to detect failures in Vision-Language-Action (VLA) manipulation by interpreting human reactions. It integrates causal paralinguistic audio detection, visual reaction recognition, stop phrases, and robot-relevance estimation. An asynchronous fusion mechanism triggers a VLA hold based on the earliest confident signal, while a separate channel handles verbal corrections. Evaluation on physical manipulation showed performance improvements through relevance estimation, demonstrating a complete pathway from social reaction to physical interruption and recovery. Latencies suggest a fast process, demonstrating a local pathway for failure detection and recovery.
Long-horizon compositional manipulation requires a rethinking of existing world-action models due to a lack of explicit subtask reasoning. Visual Goal-conditioned Action Reasoning (ViGAR) is proposed as a hierarchical framework to address this by separating planning and execution. The framework uses a visual subgoal planner to predict subgoals and a subgoal executor to generate future trajectories and actions. Both parts share a pretrained world-model, allowing for task-level planning and action generation based on common physical knowledge. ViGAR supports in-context learning, enabling different subtask decompositions by using global goal images without parameter updates, showing strong performance on benchmarks.
A model named PEWAM generates effect, robot execution, action, and terminal state streams from robot demonstrations by conditioning on the observed object movements. This model transforms demonstration data into a program of object trajectories, grasp points, and final configurations. This program can be edited by a person to specify new placements and grasp points, and it can be executed on multiple robot arms without retraining. The framework allows for re-solving closed-loop tasks from live scenes and demonstrates strong performance on held-out goals. Programs compiled from human videos have shown competitive results against methods conditioned on goal images.
MoRoOp is a dataset documenting the operations of Autonomous Mobile Robots (AMRs) in a laboratory kit preparation and supply scenario over nine eight-hour shifts. The dataset includes job specifications, operation details, dispatch events, and robot-state observations such as position, velocity, and battery charge. It contains 1,382 jobs, 4,815 operations, 19,352 dispatch events, and 140,386 robot-state observations. The data includes technically valid records of delays, navigation issues, and unsuccessful operations. This dataset is intended for evaluating AI agents on tasks like prediction, disturbance detection, and simulation.
OpenRUA introduces a zero-abstraction harness that allows coding agents to act as zero-shot visuomotor policies by accessing robot software via ROS 2. This harness minimizes bespoke abstraction layers, providing coding agents with only ROS 2 documentation and basic tools. By recasting perception as file I/O and manipulation as coding, OpenRUA demonstrated success rates of 99.0% on CaP-Bench. Coding agents spontaneously developed perception programs and motion-control clients from raw sensory inputs. This work shows that off-the-shelf coding agents can achieve robot control without specialized training or custom primitives.