Reading Modes
Papers & Evaluations with links to supporting source material.
751 selected reportsLast 30 days 660 recordsAll coverage 2,829 records
This research presents Long-WAM, a model-system framework that scales the context of causal world-action models under real-time control constraints. By using autoregressive pretraining, the model improves task success rates on benchmarks like RoboCasa GR-1 and LIBERO-Long. It also supports real-time deployment on hardware platforms, enabling dynamic and long-horizon manipulation tasks.
This paper introduces MagCilia, a compact tactile sensor that integrates a flexible magnetic-cilia structure with a Hall sensor to enable three-dimensional force sensing. Using quasi-static finite element analysis and causal history fusion regression (CHFR), the system reconstructs forces from coupled magnetic channels with root-mean-square errors of 0.40 N, 0.57 N, and 0.69 N for Fx, Fy, and Fz respectively, achieving coefficients of determination between 0.90 and 0.93. Robotic experiments demonstrate the sensor's ability to guide gripper adjustments based on tangential forces and monitor multi-axis loads under external perturbations. Additionally, frequency-domain features allow the system to distinguish six surface categories with 99.39% accuracy in cross-validation grouped by acquisition session, and an online demonstration successfully identifies all six tested surfaces.
TouchScale is a 500-hour human vision-touch dataset collected with unified sensors and protocols, enabling scaling benefits in perception and robotic manipulation. The dataset improves cross-sensor generalization and robot task success rates without using action labels, demonstrating the potential of large-scale human touch data for embodied AI.
TouchScale is a 500-hour human visual-tactile dataset collected with unified sensors and synchronization, showing scaling benefits in both perception and robotic manipulation. The dataset enables cross-sensor generalization and improves robot task success rates from 22.5% to 57.5% through mid-training. The study highlights the potential of large-scale visual-tactile data to enhance embodied AI systems.
This paper presents OpenWAM, an open framework for world-action models (WAM) that aims to standardize the evaluation of different WAM designs. The framework includes a three-stage training process, a shared MoT architecture, and four configurable interaction modes. It introduces modular components like IDM and FDM, trained on a large-scale 'counterfactual' dataset, leading to improved task success rates in both simulation and real-world robotic tasks.
Kyvern is an Apache-2.0 Python library that records a robot's decisions in a signed, hash-linked chain, allowing external verification. It supports ROS 2 with an example safety controller and provides tools for generating compliance reports. While not yet a certified safety system, it offers a framework for accountability and transparency in autonomous systems.
The European TALOS project has completed three years of testing autonomous robots and AI for solar power plant maintenance, demonstrating the technology at sites in Spain, Portugal, and the Netherlands. The system monitors solar plants, identifies performance issues, and recommends maintenance tasks to human operators. It uses production, weather, and field data alongside a digital twin to prioritize tasks. Human oversight remains central, with plant managers approving missions and adjusting parameters. The project reports potential reductions in human risk, maintenance windows, and operational costs, while citing regulatory challenges as a barrier to wider deployment.
Odyssey, an AI research company, has released Odyssey-3, its most powerful world model to date. This model enables intelligent systems to understand, predict, and interact with both physical and virtual environments. It has been adopted by companies like Flexion and demonstrated across six domains, including robotics, humanoids, vehicles, AI training, drones, and games.
TouchScale is a 500-hour human vision-tactile dataset collected with unified sensors and synchronization, showing scaling benefits in both perception and robotic manipulation. The dataset enables cross-sensor generalization and improves robot task success rates without action labels, suggesting a new path for embodied AI training.
A research team from Aarhus University in Denmark has developed a new type of tactile sensor that operates based on ionic rather than electronic signal transmission. The sensor mimics the ionic signal mechanism of human mechanoreceptors, generating voltage signals through the flow of NaCl solution in microchannels. The study demonstrates its potential applications in prosthetics and neural interfaces, and validates its frequency and sensitivity characteristics in detecting mechanical stimuli.
This paper presents PhysEvo, a framework for physical recursive self-improvement (RSI) that allows a frozen model to adapt and improve through task execution and meta-agent feedback. The framework enables joint-level control, evidence-seeking observation, and reusable manipulation skills without model-weight updates or a separately trained action policy. Evaluated on 42 RoboDojo tasks and 8 manipulation tasks, PhysEvo achieves an average score of 68.14/100 and 62.00% success, significantly outperforming existing methods. Deployment on real-world tasks shows 84.00% success across 25 trials.
This tutorial details the progression from imitation learning to Vision-Language-Action (VLA) models in robotics, focusing on the shift from repeating motions to understanding and executing tasks. It outlines key experiments with models like SmolVLA and π0.5, and discusses challenges such as data quality, model capacity, and simulation-to-real transfer, ultimately aiming to build robots that can see, understand, adapt, and act.
This article details the progression of a robotics team's work from simple imitation learning to developing Vision-Language-Action (VLA) models. It outlines challenges in adapting robotic behavior to new environments, the development of models like SmolVLA and π0.5, and the shift toward simulation-generated data for training. The focus is on building robots that can see, understand, and act autonomously in dynamic settings.
This paper describes the development of intelligent robots for assembling GB300 tester trays, a complex manufacturing task requiring high precision and adaptability. The authors explore challenges such as uncertainty in part geometry and performance requirements, and present a technical strategy combining classical robotics, imitation learning, and simulation-based reinforcement learning. The work emphasizes the importance of real-world data and infrastructure for achieving reliable automation in industrial settings.
Researchers have developed an aquatic soft robot inspired by feather stars that can move in all three dimensions using only two actuators. The design utilizes mechanical intelligence to achieve efficient 3D maneuverability through three modes: jellyfish, fish, and rotor. The robot can explore underwater spaces, lift objects, and work collaboratively with other robots, with future plans to develop a fully wireless version.
This research paper introduces a robot exposure index to assess how well robots can perform job tasks today. It finds that robots can handle 74% of physical tasks in the US, making up 34% of working hours, but are cost-competitive for only 0.3% of tasks. The study highlights barriers to adoption, including environmental adaptability and human preferences, and suggests that physical work may be automated first where robots already have a foothold.
This paper identifies modality shortcuts in vision-language-action (VLA) policies, where policies rely on superficial cues like visual, lexical, or motor signals rather than task-relevant evidence. The authors propose Perturbot, a method that introduces task-preserving wrist-view perturbations, enriched instructions with decision-relevant captions, and relabeled trajectory segments to reduce reliance on these shortcuts. They also introduce GroundFscore, an offline score to evaluate how much a policy depends on shortcuts. The work provides a framework for training and evaluating VLA policies that disentangle decisions from shortcut priors while preserving responsiveness to task-relevant evidence.
MIT researchers have developed a new autonomous navigation system called SANDO that guarantees collision-free flight for UAVs in dynamic, unknown environments. The system uses a time-sensitive safety corridor and mathematical proofs to ensure safety, even when obstacles are moving unpredictably. SANDO has been tested in simulations and real-world flights, successfully avoiding collisions in all scenarios.
This paper presents a collection of recent research papers and datasets focused on vision-language-action (VLA) models in robotics. Topics include recursive video in-context learning, simulation-guided world models, closed-loop benchmarks for driving, and methods for improving generalization and safety in VLA systems. The work highlights challenges such as task success vs. instruction following, the grounding gap, and the need for efficient and reliable adaptation in real-world scenarios.
NVIDIA has released a policy model trained on LIBERO-10, part of the Cosmos Cookbook recipe. The model, trained with 379 real demonstrations and 1137 duplicates, achieves 94.2% success in standard scenes and 85.3% in perturbed scenes. The model is available under the NVIDIA Open Model License.
Article dates follow your selected timezone. Briefing editions use Hong Kong time (UTC+8).