AI infrastructure companies are aggressively entering robotics by acquiring labs and building development platforms. This movement goes beyond simply supplying GPUs to address the complex challenges of physical AI. Companies like AMD, NVIDIA, and CoreWeave are merging their capabilities to develop intelligence that can execute physical tasks. These entities benefit by leveraging their infrastructure to build and validate physical AI models. Ultimately, this collaboration seeks to establish a new flow where AI builds AI through practical, real-world experimentation.
NVIDIA's physical AI strategy involves a three-computer system: DGX for training, Omniverse for simulation, and Jetson/IGX for real-world robot deployment. Simulation tools like Isaac Sim and Isaac Lab allow developers to train robot policies safely by learning physics through trial and error in virtual environments. The core challenge is integrating real-world physics fidelity with the need for diverse, synthetic data to ensure robots achieve robust, real-world skills. NVIDIA is working on a physics interoperability layer and unified data representations to standardize fragmented robotics development. Ultimately, the focus is on creating a platform where simulation, data, and physical deployment work together for autonomous robotics.
Eni and Generative Bionics signed an MoU to explore industrial applications for humanoid robots, focusing on inspection, teleoperation, and remote assistance. The collaboration will investigate developing lighter, stronger materials for humanoid feet and footwear, potentially leveraging Eni's chemical expertise. Cooperation also includes evaluating the use of Eni's industrial sites for robotics production and assessing access to computing resources for Physical AI development. Specific goals involve improving the impact absorption, grip, and durability of the humanoid systems. No immediate robot orders or deployment timelines were established by the agreement.
Agile Robots partnered with Hitachi to develop and commercialize Physical AI for manufacturing. This collaboration integrates Agile Robots' hardware and AgileCore platform with Hitachi's AI and industrial integration expertise. Hitachi will contribute its edge AI semiconductor technology to process visual and tactile information rapidly. The resulting solutions will be offered globally through Hitachi's HMAX Industry portfolio. The partnership aims to transform industry by reliably applying Physical AI in factory settings.
Singapore's Home Team Science and Technology Agency has opened the Home Team Humanoid Robotics Centre (H2RC) to develop humanoid robotics for public-safety missions. The center is training humanoids for tasks such as hazardous-material work, post-blast investigation, and searching prison cells. The facility supports training through realistic simulations and data collection to build vision-language-action models. The ultimate goal is for a first humanoid to work alongside officers in a real operational setting by 2028. This development involves partnerships with entities like ST Engineering and Mistral AI to ensure safe and reliable deployment.
Military humanoid robotics development involves various entities like the U.S. Army competition, Vision One's Titan, and Si Robotics' tracked alternatives. Key milestones include the xTechHumanoid competition winners and ongoing research into adaptation after deployment. Different approaches exist, such as Vision One focusing on supervised autonomy, Si Robotics on tracked mobility, and Skylark Labs on continual field adaptation. Claims of military application require evidence of operational use rather than just research or competition results. The distinction between research, proposed missions, and actual deployment is crucial for assessing progress.
The research addresses whole-body grasping and pick-and-throw for soft continuum robots using outcome-based actuation-space optimization. The method quantifies grasping by tip angular sweep and body-object enclosure, and throwing by release-direction alignment and minimum release speed. A derivative-free CMA-ES framework is employed to optimize this process, learning from sampled conditions and refining solutions. Simulation results showed high success rates, while hardware experiments achieved 100% grasping and pick-and-throw success. These results validate the framework for whole-body manipulation in both simulated and physical settings.
Fiatlux is a benchmark designed to evaluate humanoid robot capabilities in a single, long-horizon task combining ladder climbing and light-bulb replacement. The task requires a humanoid to position a ladder, climb it, exchange a light bulb in a socket, and dispose of the spent bulb safely. The episode is broken down into twelve subtasks assessed using difficulty-weighted gates to measure success in replacement, disposal, and fragility constraints. The evaluation considers both standard observations and privileged simulator states, referencing various model and controller baselines. The benchmark code and teleoperated recordings are publicly available for further use.
SynIL is a novel framework designed for automated, label-free demonstration quality assessment in offline reinforcement learning. It leverages neuroscientific principles by quantifying motor synergy, a low-dimensional movement structure, to assess motor proficiency. The method uses self-supervised reward regression to generate dense, transition-level reward signals from demonstrations. Evaluations on locomotion and manipulation benchmarks show that synergy-derived rewards strongly correlate with ground-truth rewards. SynIL's performance surpasses Behavior Cloning and matches or exceeds offline RL trained with true environment rewards.
The research introduces a two-echelon covering tour vehicle routing problem (2E-CTVRP) for post-disaster supply distribution using trucks and drones. The first echelon involves trucks and drones moving supplies from a depot to peripheral satellites, while the second echelon involves drones delivering supplies from satellites to victim clusters. The problem formulation minimizes the total arrival times of trucks at satellites and the depot. A hybrid metaheuristic, GRASP-ILS-PR, is proposed to solve this problem effectively, achieving optimal solutions quickly on benchmark instances. The analysis suggests that smaller trucks often lead to more equitable and earlier delivery times to victims.
GestAdapt is a framework designed to generate co-speech gestures for humanoid robots conditioned on the available workspace. It learns from six co-speech corpora using a shared motion representation to support retargeting across different robot embodiments. Quantitative evaluation showed that gestures generated under workspace constraints maintained high quality and ranked better than unconstrained methods. The framework suggests adapting gestures within the workspace during generation rather than post-hoc adjustments to trajectories. This approach maintains gesture naturalness while respecting physical workspace limitations.
Training data variation, such as domain randomization, is used to improve the robustness of robotic manipulation policies, but its learning mechanisms are unclear. The study investigates these mechanisms by randomizing object properties and scene variations across various robotic tasks. The empirical neural tangent kernel (NTK) is used to diagnose shifts in learning from memorization to adaptation based on the level of variation. An NTK-based signal-to-noise ratio also helps detect when policies learn to ignore irrelevant factors. These diagnostics offer practical guidance for designing training schemes and detecting shortcut learning in robotics.
PneuTac is a unified framework for tactile-feedback manipulation of soft pneumatic robots using simulation. It employs the Material Point Method (MPM) to model the robot and tactile membrane dynamics, and 3D Gaussian Splatting (3DGS) for rendering. The framework integrates real-to-sim modeling via vision and trains action/perception networks using surrogate models for efficient simulation. Experiments demonstrate that policies trained with simulation-augmented demonstrations outperform baselines in real-world tactile manipulation tasks. This provides a practical method for tactile manipulation on compliant hardware.
The research proposes a reachability-based safety certificate for motion policies in dynamical systems to ensure safety in unknown environments. This certificate is derived from a value function based on the backward reachability tube, which measures worst-case safety along a nominal trajectory. The work demonstrates that the reachability problem simplifies under specific conditions, yielding a well-defined value function that represents the maximal forward-invariant subset of the obstacle-free region. The certificate is validated across five different dynamical system constructions and tested on a Franka manipulator. Furthermore, the method shows that Control Barrier Functions (CBFs) avoid issues like saddle points that can occur in modulation strategies near unsafe zones.