Reading Modes
Open Artifacts with links to supporting source material.
782 selected reportsLast 30 days 706 recordsAll coverage 2,428 records
The article explores the integration of AI agents with robotics through edge computing, emphasizing the use of standardized protocols like MCP and the separation of cognitive and motor functions. It discusses the advantages of edge processing over cloud-based solutions, including reduced latency and cost predictability, and outlines a phased approach for deploying physical AI in industrial settings.
This article explores the evolution of embodied intelligence through world models。
CodeActionBench is a study that evaluates the performance of general multimodal agents in robot operation tasks, testing the strategy generation and execution capabilities of multiple models through a dual-arm simulation environment. The study highlights Astra's advantages in task success rate and code error rate, and analyzes differences in task coverage, execution trajectories, and completion judgment across different configurations.
In October 2026, 15 humanoid robots from Leju Robotics and Zhao Feng Machinery began operating in a car parts production line in Zhejiang, marking the first replicable embodiment of intelligent training in real-world scenarios. The robots perform a sequence of 10+ actions with 98% success rate, collecting over 40,000 data points monthly. The project includes open-source contributions to the OpenLET community, aiming to lower industry R&D barriers and enable scalable deployment.
This research presents LEGS, a hybrid simulator that combines a mesh foreground with a photorealistic 3D Gaussian Splatting background to train vision-language-action (VLA) policies for humanoid loco-manipulation without human teleoperation. The method uses a procedural motion-primitive generator to synthesize labeled demonstrations and a two-stage color calibration to align renders with the deployment camera. Policies trained on LEGS data match or exceed those trained on human teleoperation demos and outperform mesh-only baselines. The approach enables cost-effective re-rendering of demonstrations for new scenes and objects.
This research explores how agents fail to maintain goal alignment during distribution shift, identifying three distinct failure modes based on where the goal is lost within the agent's internal processing. The study shows that recovery methods vary depending on the depth of goal loss and demonstrates how steering and downstream wiring can improve goal reach in different scenarios.
This paper introduces DiffGIPC, a GPU-accelerated differentiable framework for inverse problems involving contact-rich interactions between deformable and nearly rigid bodies.
This paper presents Gaussian-LIC2, a photo-realistic LiDAR-Inertial-Camera Gaussian Splatting SLAM system that simultaneously addresses visual quality, geometric accuracy, and real-time performance. The system incrementally reconstructs a large-scale 3D Gaussian map using camera and LiDAR data, enabling high-quality novel view rendering. It employs a lightweight zero-shot depth model to address under-reconstruction in LiDAR-blind areas and uses sparse LiDAR data to supervise Gaussian map optimization. The system also improves odometry robustness through photometric constraints in a continuous-time factor graph optimization framework.
LeRobot@LeRobotHFSignalEditorial score6868
LeRobot@LeRobotHFSignalEditorial score6565CylindTrack is a depth-aware cylindrical tracking-by-detection framework for panoramic multi-object tracking in mobile robots. It addresses coordinate discontinuities and depth estimation fluctuations by jointly modeling panoramic topology and temporal depth consistency. The framework improves identity consistency and trajectory continuity, achieving HOTA scores of 33.67 and 31.12, and IDF1 scores of 40.45 and 34.33 on QuadTrack and JRDB datasets, respectively. It runs at 28.56 and 21.34 frames per second on an NVIDIA GeForce RTX 3090 GPU.
This paper introduces ASENA, an embodied agent system that enables general-purpose coding agents to program humanoid navigation, manipulation, and interaction. ASENA-VLN, a 4B monocular navigation policy, achieves state-of-the-art success rates on R2R and RxR benchmarks. Integration with coding agents improves success rates and reduces execution time, while persistent workspace evolution enhances performance through simulator feedback.
RoboGaze is a training-free multi-agent VLM evaluator that provides temporally localized diagnostic reports for robot world models. It outperforms existing baselines, including an agentic detector, and improves accuracy in detecting failures in synthetic manipulation videos, achieving 85% mean clean-clip accuracy on the strongest backbone.
This paper presents Approximate Imitation Learning, a novel framework that reduces the computational cost of training event-based control policies for quadrotors in cluttered environments. By separating representation learning from policy search, the method achieves a 28x speedup in training time and enables real-world validation with high-speed flight.
LLM2R is a new method for 3D part assembly that uses a pretrained decoder-only large language model (LLM) to predict the rigid SE(3) transformations of parts. It encodes parts into geometric tokens using a Point-BERT encoder and a Q-Former, then processes the sequence with a causal LLM adapted with LoRA. A learnable summary token block aggregates the entire assembly, and a part-aware pose head predicts rotation and translation. LLM2R achieves the highest part accuracy and success rate on PartNet categories, the best rotation and translation errors on PartNeXt, and the lowest translation error and fastest inference on Breaking Bad.
This research presents Thinker-WM, a unified world model that jointly models RGB, metric depth, and robot action through a shared intermediate representation. The model supports multiple functions including video-action generation, world prediction, and policy inference, and achieves top performance on several embodied intelligence benchmarks.
This research presents RoboFQE, an off-policy evaluation framework that estimates the discounted expected return of target robot policies using offline trajectories without requiring new rollouts or behavior-policy density estimation.
This research presents PACT, a method that jointly optimizes the mechanical design and posture retargeting of a 17-servo robot hand to better replicate human grasping postures. It also achieves an 86% grasping success rate on YCB objects, with an imitation-learning policy achieving 75% success in a pick-and-place task.
This research presents CoSMO, a framework that transforms vision-language model embeddings into online-calibrated coverage sets and safety-aware obstacle dilations to ensure safe navigation in open-world environments. By calibrating directly in the embedding space, CoSMO avoids score rescaling issues as vocabulary expands. Experiments on the Matterport3D dataset show a 61.6% reduction in semantic collisions compared to geometry-only baselines while maintaining formal coverage guarantees. Real-world tests on a quadruped robot achieved an empirical coverage of 90.73% against a 90% target, demonstrating generalization under distribution shift and delayed expert feedback.
This research proposes a dual-expert architecture, DualExpert, combining a Temporal Convolutional Network (TCN) and a Finger Graph Filter Network (FGFN), to classify grasped objects and predict finger-specific stop events using only joint-angle and actuator-load feedback. The system achieves 93.78% trajectory-level classification accuracy and 95.25% F1 score for stop prediction, with low inference latency suitable for edge deployment.
Article dates follow your selected timezone. Briefing editions use Hong Kong time (UTC+8).