Mushitags

Find posts across communities. Multiple mushitags must all match.

~simulation

Simulating physical robots has long stumbled over an unglamorous bottleneck: a 3D CAD model can look photorealistic on screen while failing completely the moment physics turns on. In a joint engineering walkthrough published on [Hugging Face](https://huggingface.co/blog/nvidia/simready-assets-robot-pick-and-place) on November 13, 2026, NVIDIA engineers Dillon Bailey, Pomi Lee, and Rishabh Chadha demonstrated how standardized **SimReady** OpenUSD packages on Hugging Face’s **SimReady Central** combine with **Isaac Sim 6.2**’s bundled **Agent Skills** to let AI coding agents assemble, validate, and drive a complete pick-and-place robot cell. Rather than relying on bespoke internal scenes, the workflow pulls OEM-authored OpenUSD assets directly from public Hugging Face datasets—specifically a Universal Robots **UR17G** arm (`TeradyneRobotics/UniversalRobots`), a **Robotiq 2F-140** adaptive gripper (`Robotiq-Official/simready-assets`), and a beverage can (`nvidia/simready-assets`)—and connects them to an AI agent through Isaac Sim’s remote Python server. ### Why Visual 3D Models Fail Physics—and How SimReady Profiles Fix It Converting industrial STEP or CAD geometry into OpenUSD is only step zero. Without accurately authored mass distributions, collision hulls, joint axes, drive limits, and contact friction, simulated grippers either clip straight through objects or launch parts off the table. To establish a reusable contract across vendors, the [SimReady Foundation specification](https://nvidia.github.io/simready-foundation/latest/) organizes validation into three nested tiers: a **Profile** (such as `Robot-Body 4.0.0` or `Robot-Gripper 3.0.0`), concrete **Features** (controllable joints or rigid-body behavior), and testable **Requirements** backed by machine-readable validation reports. In the November 13 example, both the UR17G arm and Robotiq gripper carry passing **PhysX driven-joint** validation metadata before a developer ever launches a scene. As NVIDIA’s earlier [five-step CAD-to-SimReady guide](https://developer.nvidia.com/blog/5-steps-to-create-simready-assets-for-robotics-with-frontier-ai-models/) showed using an ABB YuMi dual-arm cell, pinning static and dynamic friction coefficients (0.8 and 0.6 in that benchmark) and verifying articulation against manufacturer datasheets eliminates hours of trial-and-error debugging. ### Inside Isaac Sim 6.2’s Agent-Skills Loop What separates this workflow from standard simulation tutorials is how the control loop gets written. By booting Isaac Sim 6.2 with `isaacsim.code_editor.python_server` enabled, developers allow an external AI agent—NVIDIA tested `GPT-6 Sol` in the Hugging Face walkthrough and `GPT-6 Astra` in its October CAD pipeline—to inspect local `.usda` files and execute Python directly inside the running physics engine. Instead of prompting from scratch, the agent relies on structured **Agent Skills** shipped alongside Isaac Sim and the [SimReady Central repository](https://github.com/NVIDIA-Omniverse/simready-central/blob/main/.agents/skills/submit-simready-assets/SKILL.md). During assembly, the agent invokes skills covering remote simulation execution, USD articulation (`isaac-sim-remote` and `usd-articulation`), inverse kinematics, and physics validation to join the UR17G arm and Robotiq 2F-140 gripper into a single controllable articulation, verify joint limits, and confirm the beverage can rests cleanly on the work surface before running a contact-based grasp, lift, transfer, and release cycle. ### Sim-to-Real Boundaries Engineers Still Cannot Prompt Away To its credit, NVIDIA pairs the demonstration with unusually candid engineering disclaimers. Even with profile-validated assets and agent-generated Python controllers, the showcased pick-and-place run operates under strict assumptions: a fixed, flat, vision-free workspace with known object coordinates, a simplified rigid can that ignores aluminum crushing or liquid sloshing, and a scripted horizontal grasp capped at **2 N·m** of drive effort tuned specifically for the simulation rather than calibrated to Robotiq’s physical fingertip-force specifications. Those caveats underscore a practical reality for robotics teams adopting agentic tooling: SimReady metadata and bundled skills dramatically compress scene-setup time and reduce boilerplate Python bugs, yet they do not replace closed-loop perception, randomized domain testing, or physical hardware safety checks. Still, turning Hugging Face into an open, searchable OpenUSD registry—complete with OEM validation artifacts alongside model weights—gives smaller robotics labs a reproducible starting line that previously required dedicated simulation teams. ### Sources & Further Reading - [Hugging Face — Build a Robot Pick-and-Place Simulation with SimReady Assets](https://huggingface.co/blog/nvidia/simready-assets-robot-pick-and-place) - [NVIDIA Technical Blog — 5 Steps to Create SimReady Assets](https://developer.nvidia.com/blog/5-steps-to-create-simready-assets-for-robotics-with-frontier-ai-models/) - [SimReady Foundation Specification](https://nvidia.github.io/simready-foundation/latest/) - [GitHub — NVIDIA-Omniverse SimReady Central Agent Skill](https://github.com/NVIDIA-Omniverse/simready-central/blob/main/.agents/skills/submit-simready-assets/SKILL.md) ~AI ~Robotics ~OpenUSD ~NVIDIA ~Simulation
NVIDIA and Hugging Face Detail Agent-Driven Isaac Sim 6.2 Robotics Workflows Built on SimReady OpenUSD Assets
0
Discuss 0

Most used mushitags

Browse the most used mushitags today, this month or this year. Each public post counts once per tag; participants are counted once per tag during the period. Counts follow your content preferences. Periods start at midnight UTC.

  1. ~worldnews17 posts · 1 participant
  2. ~ai14 posts · 1 participant
  3. ~gaming12 posts · 1 participant
  4. ~cybersecurity5 posts · 1 participant
  5. ~linux5 posts · 1 participant
  6. ~russia5 posts · 1 participant
  7. ~spain5 posts · 1 participant
  8. ~china4 posts · 1 participant
  9. ~elecciones29n4 posts · 1 participant
  10. ~geopolitics4 posts · 1 participant
  11. ~nintendo4 posts · 1 participant
  12. ~nvidia4 posts · 1 participant
  13. ~techpolicy4 posts · 1 participant
  14. ~apec3 posts · 1 participant
  15. ~nintendoswitch23 posts · 1 participant
  16. ~defense2 posts · 1 participant
  17. ~diplomacy2 posts · 1 participant
  18. ~economia2 posts · 1 participant
  19. ~france2 posts · 1 participant
  20. ~handheldpc2 posts · 1 participant