Reading Modes
Open Artifacts with links to supporting source material.
630 selected reportsLast 30 days 568 recordsAll coverage 2,235 records
This research paper introduces Agi, a large-scale dataset with over 1 million trajectories across 217 tasks, and presents Genie Operator-1, a policy that achieves 30% performance improvement over existing methods. The dataset and tools are open-sourced to advance research in scalable embodied intelligence.
This work proposes ASTEROID, a student-teacher supervised learning framework that enables agents to learn efficient test-time exploration under partial observability. By distilling expert actions conditioned on the student's own on-policy history, the method resolves uncertainty and exploits revealed states without requiring privileged expert exploration or impractical reinforcement learning training costs. Across eleven diverse environments—including robotic manipulation, navigation, and discrete benchmarks—ASTEROID achieves approximately 2x higher returns under fixed exploration budgets and uses roughly 100x fewer interactions than RL baselines. The approach also transfers from simulation to vision-denied real-world robot manipulation, increasing task success by +40% over standard distillation.
PlanSys2 4.0.0, a PDDL planning and execution framework, has been released with enhanced reliability and robustness. It supports PDDL from version 1.2 to PDDL+, including advanced features like numeric fluents, durative actions, and PDDL+ processes. The release includes the Kairos planner as the default, which offers broader PDDL coverage and faster performance. The update is available for multiple ROS distributions.
The author presents an experiment using C++26's std::execution model to decouple computation from ROS 2 callbacks, enabling more explicit control over execution resources. This approach allows for better management of computational tasks in scalable robotic systems, particularly in drone swarms, by separating expensive operations from communication and coordination callbacks.
Li Rongzhong's Petoi company has successfully transitioned from a university project to a commercial product line through three generations of quadruped robot products. The article details the technical challenges, cost control, user feedback, and market strategies during the product iteration process, while also exploring the complexities and practical difficulties of applying robotic products in real-world scenarios.
Hugging Face and NVIDIA have released SimReady Central, an open-source library of simulation-ready 3D assets for robotics and AI factories. The platform provides over 2,000 OpenUSD assets with pre-configured physical properties, enabling developers to build training environments and digital twins more efficiently. It supports a variety of use cases, including robot manipulation and AI factory simulations, and allows OEMs to contribute and validate assets against SimReady standards.
NVIDIA announced at GTC Berlin the expansion of its Isaac open robot development platform, including a humanoid robot reference platform, updated tools, and datasets. The Agile ONE SN, set for early 2027, integrates simulation, training, and deployment tools, with hardware and software designed for rapid development and real-time operation.
NVIDIA announced at GTC Berlin the expansion of its Isaac platform with the Agile ONE SN reference humanoid robot, new datasets, and runtime tools. The system integrates simulation, training, and deployment tools to shorten setup time and enable real-world deployment. Agile ONE SN features a 1.40-meter humanoid with 31 degrees of freedom, WUJI Hand 2 hands, and Jetson Thor for real-time control and perception.
A Chinese company affiliated with Tsinghua University has released a prediction-based world model called Polibrain, designed to enable robots to predict and plan actions in dynamic environments. The model focuses on state representation and generalization across different robot morphologies, with a theoretical foundation in polynomial representation and causal memory. The research is positioned as a contribution to the broader field of embodied AI and world modeling, with potential applications in various robotic tasks.
Bear Robotics has launched KINO-2, a robot-learning platform that transforms task demonstrations into deployable manipulation skills for its KR1 humanoid robot. The platform integrates with NVIDIA Isaac OS and leverages foundation models to enhance inference speed and operational efficiency in commercial environments.
This research presents an end-to-end pipeline for training a surgical robot policy using motion data captured from standard laparoscopic instruments. The system uses inertial, time-of-flight, and Hall sensors to track instrument state without external cameras. A diffusion policy is trained on both ex-vivo and in-vivo demonstrations, achieving appendectomy in three of four live rabbits. The robot serves as a timing reference and executor, with no demonstration collection.
This paper presents LocoPilot, a single-operator bilateral teleoperation framework for loco-manipulation policy learning. It enables direct capture of robot-specific signals like proprioception and body-mounted observations without motion-capture instrumentation. The system integrates a 12-DoF quadruped with a 6-DoF manipulator, using two handheld controllers and a follower arm for force feedback. A ROS~2 pipeline records synchronized multimodal data and a unified 11D action space, validated through deployment on a physical robot without retargeting or additional inverse kinematics.
This paper introduces EchoNav, a framework that combines temporal aggregation and occupancy completion in an evidential bird's-eye-view (BEV) representation to enable visual navigation in cluttered environments. The system adapts frozen VGGT-Omega features using historical residual proposals with cellwise weights and a learned no-update option, producing local maps without depth or camera-pose inputs. Trained on simulated RGB sequences with visibility-masked occupancy annotations, EchoNav achieves 86.3% success in simulated two-goal navigation (outperforming the strongest baseline by 32.2 percentage points) and 73.3% success across 60 trials on a wheeled robot in 20 scenes, compared to 33.3% for the baseline. Additional demonstrations were conducted on quadruped and humanoid robots.
This research paper presents a runtime safety check for robot control loops, showing how a model can execute harmful instructions without refusal. The method stops harmful actions by asking the model to evaluate the instruction before execution, preventing unsafe behavior in robotic systems.
OpenAMRobot 2.0 is a general-purpose, open-source dual-arm mobile manipulator with an embodied AI software layer, built on ROS 2. It features two Enactic OpenArm 2.0 arms, a NVIDIA Jetson Orin NX compute module, and a safety layer with emergency stops. The robot is designed for tasks like grocery order picking and is being tested in Berlin, with a final release planned for December 2026.
NVIDIA has released 'Flora', a suite of open AI models designed to automatically generate robot learning data. Flora includes three models for action segmentation, captioning, and reward generation, and is claimed to outperform other models in certain benchmarks while being significantly more cost-effective for large-scale video processing.
NVIDIA has released Cosmos3-H-Embodied-Nano, a multimodal model trained on surgical robotics data for forward dynamics, inverse dynamics, and policy prediction. It is based on the Cosmos3-Nano model and is intended for use in surgical world modeling and imitation learning.
Ropedia showcased six research projects at Ropedia Super Week, including two Foundry initiatives (EgoXperience and HoliWorld) for structuring egocentric human experience into data, and four studies on spatial reasoning, world modeling, and robot learning and control. The projects include datasets like EgoX-Data-6M with approximately six million frames, benchmarks showing performance improvements such as a 13.9-point score increase on MMSI-Bench for S-Agent 2.0, and simulation results indicating task success rates of 55% for model-guided action selection versus 26% for random selection. Ropedia aims to build data infrastructure connecting human experience with robot capabilities for Physical AI.
NVIDIA has released a trained policy model for the Cosmos3-Nano robot, post-trained on the LIBERO-10 environment. The model includes training checkpoints and evaluation metrics for both standard and perturbed scenes, with success rates of 93.3% and 85.5% respectively. The model is part of a synthetic-data study and is licensed under the NVIDIA Open Model License.
NVIDIA has released a trained policy model for the Cosmos3-Nano robot, post-trained on the LIBERO-10 environment. The model, based on training seed 43, achieves 94.9% success in standard scenes and 89.7% in perturbed scenes. The model is part of a synthetic-data study and is available via the cosmos-framework policy server.
Article dates follow your selected timezone. Briefing editions use Hong Kong time (UTC+8).