Newton's Classroom - The Physics Class Every Robot Should Attend скачать в хорошем качестве

Newton's Classroom - The Physics Class Every Robot Should Attend

Newton Physics Engine represents a breakthrough in robotics simulation, offering an open-source, GPU-accelerated platform that bridges the critical gap between simulated and real-world robot performance. Developed collaboratively by NVIDIA, Google DeepMind, and Disney Research, and managed by the Linux Foundation, Newton is built on NVIDIA Warp and OpenUSD frameworks to provide high-fidelity physics simulation with unprecedented speed and accuracy.​ In this livestream, we'll explore Newton's modular architecture supporting multiple physics domains—rigid body dynamics, soft body simulation, cloth modeling, and granular materials. You'll discover how Newton's differentiable physics enables gradient-based optimization for efficient robot learning, and how it can be used with robot learning frameworks like NVIDIA Isaac Lab and MuJoCo playground. Key Moments: 00:39 – What Newton is: open-source, GPU-accelerated physics (NVIDIA + DeepMind + Disney; Linux Foundation; OpenUSD + Warp) 03:28 – Why Newton: robotics spectrum & stack complexity it’s built to tackle 07:05 – Data strategy: real ↔ synthetic ↔ internet-scale; why synthetic matters 10:50 – Design requirements: modularity, extensibility, differentiability, unified API 16:03 – Architecture & integrations: solvers, collision, models; Isaac Lab now, Isaac Sim planned 25:20 – NVIDIA Warp refresher: Python DSL → CUDA, autodiff, PyTorch/JAX interop 43:03 – Solver lineup: MuJoCo-Warp, XPBD, Featherstone, VBD cloth, implicit MPM, Camino (Disney), neural solver 47:34 – Benchmarks: MuJoCo-Warp vs MJX—large throughput gains on RTX GPUs 50:12 – Deformables & mixed systems: cloth + rigid, granular sand; 1-way vs 2-way coupling 1:10:13 – Roadmap: Beta → Beta-2 polish → GA target; Camino on the path to GA Q&A Highlights: 04:03 — Q: What robot-simulation challenges does Newton aim to resolve? A: Unifying high-fidelity physics under a modular, open, GPU-accelerated framework with differentiability and a simple, unified API so you can mix solvers (rigid, soft, cloth, granular) and scale synthetic data for learning and control. 16:24 — Q: How does the unified API expose gradients to outside frameworks (e.g., MuJoCo / DL libs)? A: Newton leverages NVIDIA Warp’s automatic differentiation. Gradients are opt-in and retrievable via Warp APIs with zero-copy array/gradient interop to PyTorch today (JAX gradient interop is planned in the next Warp release). Newton inherits this through Warp. 19:06 — Q: Will DGX Spark run Isaac Sim and Isaac Lab? A: Yes for Isaac Lab—an experimental branch with Newton ran on Spark; standard Lab runs on Spark as well. Isaac Sim + Newton is planned as a separate backend (timeline in progress). 31:23 — Q: How are state/collision unified (OpenUSD + Warp) to switch solvers during one sim? A: Import assets via USD/MJCF/URDF, build a Newton model, then choose the solver per model via the Newton API. A Newton USD schema is being developed to make this smoother; for now, model build + solver selection handles the handoff. 33:50 — Q: Can Newton run as an Omniverse Kit/AppKit extension? A: Newton is usable standalone and in Omniverse contexts. For Isaac Lab, Newton is already integrated (experimental). For Isaac Sim, Newton is coming as an additional physics backend (default status TBD). Warp already has a Kit extension. 37:47 — Q: Does Warp’s autodiff enable end-to-end training of physics-based controllers in-sim? A: Yes—that’s the idea. With differentiability you can compute gradients inside the sim loop and feed them to your training/optimization (e.g., PyTorch), enabling online/in-place learning and control optimization. Resources: [Talk] Announcing Mujoco-Warp and Newton: How Google DeepMind and NVIDIA are Supercharging Robotics Development (https://www.nvidia.com/en-us/on-deman...) [Talk] How Robots Learn To Be Robots in Simulation With the Newton Physics Engine (https://www.nvidia.com/en-us/on-deman...) [Talk] How Disney Droids Come to Life With Physics Simulation (   • How Disney Droids Come to Life With Physic...  ) [Whitepaper] Isaac Lab: A GPU Accelerated Simulation Framework For Multi-Modal Robot Learning (https://research.nvidia.com/publicati...) Got questions, post them on our Discord Thread:   / discord   Discord Invite:   / discord   📆 Check out the full calendar for all of our upcoming events → https://nvda.ws/3JqaWnA ---------------------------------------------------------------------------- ⬇️Get Started → https://nvda.ws/4cZAZO1 ​ 👀Explore OpenUSD → https://nvda.ws/3CeozBQ ​ 👥Join the Community → https://nvda.ws/3ZMfc6e