The robotics industry's reliance on teleoperation, where humans demonstrate tasks remotely, is a flawed approach that creates a structural wall due to the vast amount of data required. Current datasets are 100,000 times smaller than those used to train language and vision models, making it impossible for operators to generate enough demonstrations to keep up with the pace of technological advancements. This approach ultimately leads to a system that cannot generalize across real environments without human intervention.
Researchers have created an "Interactive World Simulator" that allows robots to be trained and evaluated on a simulated environment, reducing the need for real-world demonstrations. The simulator uses an action-conditioned video prediction model to generate realistic images and can run interactively for over 10 minutes at 15 FPS. This technology has the potential to unlock two long-standing bottlenecks in robot learning: cheap data generation and scalable policy evaluation.
Xiaomi-Robotics-1, a vision-language-action (VLA) model, achieves strong scaling behavior and outperforms state-of-the-art methods in multiple simulation benchmarks. The model is pre-trained on over 100k hours of real-world manipulation trajectories collected via UMI devices and fine-tuned for complex, dexterous tasks with high data efficiency. It establishes a new state-of-the-art with a 57.6% success rate on RoboCasa365 and outperforms the prior state-of-the-art by 11.1%. Code and model checkpoints will be released.
NeuroCommitSSM, a decision-centric framework, predicts continuous commit-readiness scores from EEG, EMG, and ET data to ensure safe assistive robotic manipulation. The three-state finite-state supervisor, HOLD-ASSIST-COMMIT (HAC), gates execution requiring both sustained neural model signals and real-time perception and robot-state feasibility. NeuroCommitSSM achieves 0.950 action-balanced accuracy with low false commit events per 1000 REST windows.
Soft robots, well-suited for rehabilitation and surgery, require reproducible fabrication to ensure real-world deployment. The authors present robust procedures for two-part pour casting of soft pneumatic actuators, preventing internal cavity clogging and air-tight sealing. A sensor embedding procedure for thin-film flex sensors is also presented, allowing accurate data acquisition. Finite Element Modeling and experiments validate the performance of the fabricated actuator, showing repeatable performance across two operators and 24 successful fabrications.
A tactile-reactive gripper integrates Visuo-Tactile Active Palm (VTAP) and compliant fingers with tactile array sensors to achieve dexterous in-hand manipulation. The design combines long-range visual localization with contact-rich tactile feedback, extending the system's manipulation capability. Extensive experiments validate the system across various tasks, including reactive grasping, fragile object handling, and vision-tactile peg-in-hole insertion.
Existing approaches to learning reward functions from human preferences assume humans use expected utility, but behavioral evidence shows humans are risk-sensitive and exhibit loss aversion. A study compares expected utility with Cumulative Prospect Theory within a preference learning framework, finding CPT-based learners recover reward functions with lower regret than EU-based learners when preferences are generated by risk-sensitive users. This highlights the importance of modeling human risk sensitivity in learning rewards from preferences over stochastic outcomes. Preliminary experiments show this effect is substantial.
Researchers formulated workspace design as an optimization problem to improve shared autonomy systems. They derived a probabilistic correctness guarantee under a bounded noise model, which showed that optimized layouts can improve goal inference reliability and reduce ambiguity. Simulation experiments across multiple tabletop scenarios demonstrated the effectiveness of this approach. A real-world shared autonomy system integrated the proposed inference framework.
This paper presents a unified algorithmic framework for multi-objective kinodynamic motion planning, addressing three problem classes: lexicographic optimization, constrained optimization, and Pareto-front approximation. It proposes the Stable Sparse-RRT (SST) algorithm with modifications to maintain a representative set of locally Pareto-optimal nodes, resulting in distinct algorithms for each problem class. The paper provides theoretical guarantees for completeness and optimality and demonstrates their effectiveness through empirical evaluations.
PACE, a novel framework, enables the interactive generation and deployment of structured personas on humanoid robots through an Interactive Persona Elicitation Pipeline. The system uses user Q&A to dynamically synthesize a tailored, psychologically grounded identity, which informs a structured persona prompt built upon multi-perspective dimensions. A comprehensive empirical HRI evaluation assesses the impact of dynamically generated personas on various user interactions and outcomes, establishing a scalable pathway for deploying personalized identities in embodied humanoid assistants.
Researchers have developed a model-based strategy to decouple proprioceptive and contact sensing from fluidic pressure sensors in architected soft robots. A piecewise constant curvature model maps pressures to shape, while Huber regression identifies outlier channels indicating external contact. The approach achieves a relative bending error of 0.11 ± 0.02 and a contact detection rate of 97% across 178 trials, suggesting localized fluidic innervation combined with redundancy resolution is a practical path toward concurrent proprioception and contact sensing in architected soft robots.
MemoGuard, a lightweight adaptive runtime, validates episodic memories against topology, resource, and outcome contracts before reuse, invoking fallback only when validation fails. This reduces battery safety violations by 76.6% over similarity-only top-1 reuse and by 21.4% over always reasoning in a graph-based corridor-inspection simulator. MemoGuard opens-source implementation is available at https URL.