A new arXiv paper introduces RoboSimStudio (RSS), an interactive toolkit for rapidly converting real robot manipulation tasks into editable simulations. It makes solver, contact model, and physical parameter choices explicit and adjustable, letting both humans and coding agents iteratively refine simulations against observed physical behavior. For practitioners, it addresses the hard problem of modeling contact mechanics, deformation, stiffness, and friction in robot-object interactions.
Researchers introduced Agentic Real2Sim, a framework that uses vision-language agents to automate the labor-intensive process of converting real-world scenes into physics simulations. The system recovers scene geometries, object states, and physical parameters, then assembles actors, objects, cameras, poses, and trajectories into a runnable simulation, eliminating manual tuning and brittle workflow glue.
Researchers introduced Agentic Real2Sim, a framework that uses vision-language agents to automate the labor-intensive process of converting real-world scenes into physics simulations. The system recovers scene geometries, object states, and physical parameters, then assembles actors, objects, cameras, poses, and trajectories into a runnable simulation, eliminating manual tuning and brittle workflow glue.
A research team presents Touch2Robot, a framework letting humans collect dexterous manipulation demonstrations while seeing how the target robot hand would contact objects. It captures human hand motion, tactile-glove measurements, and object motion, then uses object-specific RL and a distilled real-time retargeter to reconstruct robot-object contacts. Across four real-world tasks it roughly doubles demonstration replay success and cuts collection time, improving downstream Diffusion Policy performance.
A research team presents Touch2Robot, a framework letting humans collect dexterous manipulation demonstrations while seeing how the target robot hand would contact objects. It captures human hand motion, tactile-glove measurements, and object motion, then uses object-specific RL and a distilled real-time retargeter to reconstruct robot-object contacts. Across four real-world tasks it roughly doubles demonstration replay success and cuts collection time, improving downstream Diffusion Policy performance.
A new arXiv paper presents ForceTwin, which builds physics-informed digital twins of articulated objects by having a person probe them with a handheld force-sensing gripper, yielding synchronized poses and interaction forces. The system estimates inertia, Coulomb friction, viscous damping, and state-dependent mechanism forces, nearly halving inertial-parameter error versus a VLM prior. Used as a feedforward dynamics model for impedance control on Spot and Franka FR3 robots, it raises goal completion to 87% across nine object-embodiment pairs, and its twins enable training whole-body door-traversal policies deployed in the real world.
A new arXiv paper presents ForceTwin, which builds physics-informed digital twins of articulated objects by having a person probe them with a handheld force-sensing gripper, yielding synchronized poses and interaction forces. The system estimates inertia, Coulomb friction, viscous damping, and state-dependent mechanism forces, nearly halving inertial-parameter error versus a VLM prior. Used as a feedforward dynamics model for impedance control on Spot and Franka FR3 robots, it raises goal completion to 87% across nine object-embodiment pairs, and its twins enable training whole-body door-traversal policies deployed in the real world.
Researchers introduce RAPID, a framework where coding agents automatically generate, verify, and refine robot programs from a single visual human demonstration. It infers task specifications, action primitives, and an interactive environment from the demo, using an object-centric relational program representation for reuse. RAPID showed strong performance in simulation on contact-rich nonprehensile manipulation tasks and LIBERO-Pro, and on a real Franka arm.
Researchers introduce RAPID, a framework where coding agents automatically generate, verify, and refine robot programs from a single visual human demonstration. It infers task specifications, action primitives, and an interactive environment from the demo, using an object-centric relational program representation for reuse. RAPID showed strong performance in simulation on contact-rich nonprehensile manipulation tasks and LIBERO-Pro, and on a real Franka arm.
An arXiv paper introduces RoboCousin, a pipeline for building custom simulation playgrounds to train robust bimanual robotic manipulation policies. It targets the data bottleneck in bimanual manipulation, where physical demonstrations are expensive and existing simulation pipelines are locked to closed asset libraries and predefined scenes.
An arXiv paper introduces RoboCousin, a pipeline for building custom simulation playgrounds to train robust bimanual robotic manipulation policies. It targets the data bottleneck in bimanual manipulation, where physical demonstrations are expensive and existing simulation pipelines are locked to closed asset libraries and predefined scenes.