NVIDIA Omniverse NuRec reconstructs real-world driving scenes using 3D Gaussian splatting to generate new camera views for target vehicle configurations. This method leverages a dataset of over 1,500 reconstructed scenes to facilitate carline adaptation by simulating views from a new vehicle rig. The process involves pairing reconstructed scenes with target rigs, rendering, refining frames with NVIDIA Harmonizer, and training a perception model on the synthetic data. NuRec enables developers to adapt perception software across different vehicle platforms without needing expensive, new real-world data collection for each carline. The workflow is further automated using agent skills to handle data downloading, rendering, and post-processing.
Kinetic Blocks launched a gated beta marketplace for humanoid training data, facilitating the buying and selling of datasets like egocentric video and robot execution recordings. The platform ensures quality through a dual-track evaluation involving file verification and a Knowledge Base Quality Score grading datasets from A+ to F. It aims to standardize procurement for physical data in the embodied AI sector, addressing the divide between video and high-friction teleoperation data. Listings are structured around chain-of-custody documentation and support formats such as LeRobot v3.0. Kinetic Blocks plans to scale operations and will seek a seed funding round in Q4 2026.
A user is seeking discussions about the implementation of humanoid robots in manufacturing facilities. They are interested in learning about use cases, safety concerns, successes, and failures related to these robots. The individual is attempting to gather information to aid in their decision-making process. Efforts to connect with robot manufacturers for discussions have been unsuccessful so far. The user is requesting any relevant information from the community.
Haptic Foundation Models (HFMs) aim to advance embodied AI by addressing the lack of generalized touch sensing in current models. This transition requires a shift from passive language/vision models to active HFMs across dimensions like action coupling and future state prediction. The research synthesizes tactile datasets and benchmarks models on tasks such as force estimation and slip detection using datasets like TacBench. This work focuses on developing capabilities for physical interaction necessary for consumer electronics and robotics. Ultimately, HFMs offer a path to more adaptive and physically aware AI systems.
CogRun is a framework enabling safety-critical ground robots to perform runtime learning on edge-AI devices in unknown environments without prior knowledge. It comprises three components: a Learning-Agent utilizing cognitive-neural learning, a Rational-Agent for safety, and a Coordinator for interaction management. The Learning-Agent employs replay buffers and safety-aware action blending from reinforcement learning and instance-based learning. This system allows for full autonomy on edge devices, bypassing reliance on wireless communication for challenging tasks. Experiments confirm CogRun facilitates safe and efficient runtime learning for robots in real and simulated wild environments.
RoboGesture is a framework designed to generate real-time, semantically-aligned co-speech gestures for humanoid robot interaction with humans. It addresses challenges in human-robot interaction by creating a system where robots listen, respond, and gesture synchronously. The method involves a Hierarchical Semantic-Acoustic Aligner to extract cues from audio and a Streaming Conditional Motion Generator based on a diffusion transformer. Anti-Inertia CFG Masking ensures proactive control signal mining from audio, and an MPC-based filter guarantees safe physical execution. Experiments on a Unitree G1 humanoid showed that RoboGesture produced safer and more semantically appropriate responses than existing methods.
RoboPhys-3D introduces a 3D-grounded Embodied World Model (EWM) benchmark built on RoboTwin 2.0 to evaluate world models. It processes generated and ground-truth videos through the same 3D reconstruction pipeline to separate reconstruction errors from generation errors. The benchmark includes 50 complementary metrics across four levels, organized into sub-dimensions like pixel fidelity and state-level understanding. Average Full Score and RoboPhyscore are introduced as hierarchical and task-aligned scores, respectively, for comprehensive evaluation. Cosmos 3 achieved the highest RoboPhyscore, showing that execution-grounded metrics are critical for evaluating EWM capability.
Automated drift controllers often fail to account for the continuous transition from grip driving to drift near the handling limit. This paper introduces a Boundary Exploration Learning Model Predictive Control (BE-LMPC) to model drift emergence without explicit drift references. The BE-LMPC iteratively explores spatial speed allocation to discover emergent drift when performance demands approach tire capacity. Analysis shows drift is a conditional continuation of limit handling, not a separate motion mode, and is dependent on external conditions. Specific results indicate drift emerges at higher friction coefficients as performance demands increase.
Indoor 3D Scene Graphs link geometric primitives to higher-level metric-semantic concepts for spatial reasoning. The proposed method is a unified autoregressive diffusion-based model that jointly learns the structure and features of these graphs bottom-up from vertical planes. This approach surpasses other learning-based and random baselines across diverse 3DSG datasets. The work also introduces an adaptation of the Fused Gromov--Wasserstein distance for evaluating generated 3DSGs against ground truth. The method is applicable to synthetic scenes, real architectural plans, and robotic sensor data with varying complexity.
Autonomous vehicle sensor fusion relies on accurate camera-LiDAR calibration, which can drift due to real-world factors. This paper introduces the Adversarial Calibration Attack (ACA), a physical attack targeting online sensor calibration. ACA spoofs the misalignment detector to force the calibration estimator toward an incorrect transformation. Evaluation shows ACA can cause significant errors, such as up to 33.9 degrees of rotational calibration error in benchmark datasets. These results highlight online calibration as a practical and safety-critical attack surface for autonomous vehicles.
The Park-Ravani construction generates a frame-invariant spline on SO(3) by exponentiating cubic canonical-coordinate polynomials. This construction transfers to orthogonal dual tensors, representing rigid displacements, through the dual extension of the right Jacobian. The method interpolates rigid poses and ensures continuity of the body dual twist and its first derivative, leading to a continuous physical acceleration field. The work distinguishes algebraic dual transfer from temporal differential prolongation and addresses holonomic defects in dual interpolants using Hermite interpolation.