RoboticsMachines in the physical world
NVIDIA Tutorial Uses an AI Coding Agent to Build an Isaac Sim Pick-and-Place Demo From SimReady Assets
An NVIDIA tutorial shows a coding agent using Isaac Sim 6.2 skills to assemble a Universal Robots UR17G arm, Robotiq 2F-140 gripper and beverage can from Hugging Face's SimReady Central library into a working pick-and-place simulation.
From the Hugging Face article of November 13, 2026 and cited sources: Most of the work is done by a coding agent. [1] [6]
SOURCES[1] Build a Robot Pick-and-Place Simulation with SimReady… huggingface.co[6] Isaac Sim 6.2 Agent Skills documentation docs.isaacsim.omniverse.nvidia.com
AI-assisted, not human-edited. Methodology Corrections Who owns this site
- Original source
- Hugging Face article, company source
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- November 13, 2026 article published
- Covered
- Nov 15 edition
- Evidence
- 6 sources across 4 sites
- 1Build a Robot Pick-and-Place Simulation with SimReady…huggingface.co
- 2Robotiq 2F-140 validation reporthuggingface.co
- 3Announcing SimReady Central (Hugging Face blog)huggingface.co
- 4NVIDIA Opens SimReady Central (Unite.AI)unite.ai
- 5Isaac Sim 6.2 General Availability (NVIDIA forum)forums.developer.nvidia.com
- 6Isaac Sim 6.2 Agent Skills documentationdocs.isaacsim.omniverse.nvidia.com
Reach: not measured yet
Most of the work is done by a coding agent. It sends Python to a live Isaac Sim 6.2 instance through the remote Python server and follows four bundled skills covering remote control, articulation, inverse kinematics and physics. The skills are version-controlled Markdown procedures and work with Cursor, Claude Code and Codex CLI. [1] [6]
Asset validation is scored separately for each physics engine. The Robotiq 2F-140 report lists 24 passed checks and 7 failed ones, including a PhysX feature check and some MuJoCo and Newton variants. Teams should read each asset's report for the engine they use before relying on it. [1] [2]
The demo is deliberately narrow. The 2 N·m grip-effort cap was chosen through simulation trials, not taken from manufacturer guidance. The arm follows fixed waypoints with no vision or replanning, and the rigid can always starts in a known position. NVIDIA says the results don't carry over to real hardware. Because the agent writes the controller code, results can vary between runs. [1]
pinned, maker-contributed assets and agent skills cut scene setup in a documented demo. Implication: integrators may prototype faster, but grip and waypoint tuning, per-engine validation review, code review and sandboxing are still on them.
Read the full assessment
The success results come only from NVIDIA.
Executive brief
The arm-plus-gripper pick-and-place in NVIDIA's new tutorial was put together and run mostly by an AI coding agent working through Isaac Sim's bundled "skills". The demo arm uses a 2 N·m grip-effort cap that NVIDIA picked through simulation trials. It is not calibrated to real fingertip forces, and NVIDIA says the results don't carry over to real hardware (NVIDIA/HF blog). The tutorial shows how to combine assets from SimReady Central, a Hugging Face library contributed by equipment makers, into a working scene. The validation reports are less tidy than the blog's "passing" summary suggests: the gripper's report lists 7 failed checks next to 24 passed ones (validation JSON). No independent evaluation was found.
What changed and event timeline
SimReady Central launches
NVIDIA opened a free, ungated library of more than 2,000 OpenUSD assets on Hugging Face. Contributors include FANUC, Universal Robots, Robotiq and Vertiv (;).
Isaac Sim 6.2 GA
NVIDIA calls it a bug-fix and stability release that supports Isaac Lab 3.0. It adds Robotiq gripper teleoperation support and better robot validation ().
Pick-and-place tutorial published
A UR17G arm, a Robotiq 2F-140 gripper and a beverage can are assembled and run in Isaac Sim 6.2, using agent-generated code (GPT-6 Sol in the example) ().
Capabilities and access
- Software: Isaac Sim 6.2, with the remote Python server enabled (
isaacsim.code_editor.python_server). - Assets: UR17G (ur17g_1300), Robotiq 2F-140, and the NVIDIA beverage can, each pinned to a specific commit.
- Agent skills used:
isaac-sim-remote,usd-articulation,manipulation-ikandphysics-simulation.
Read the full section
- Software: Isaac Sim 6.2, with the remote Python server enabled (
isaacsim.code_editor.python_server). Assets download through the website or the Hugging Face CLI (blog). - Assets: UR17G (ur17g_1300), Robotiq 2F-140, and the NVIDIA beverage can, each pinned to a specific commit. The blog's "Robotiq 2F-140" link actually points to a 2F-85 file.
- Agent skills used:
isaac-sim-remote,usd-articulation,manipulation-ikandphysics-simulation. The skills documentation says they work with Cursor, Claude Code and Codex CLI (Agent Skills docs).
Technical analysis for researchers and developers
- Spec hierarchy: SimReady Foundation sets requirements in three layers. The arm is validated against Robot-Body 4.0.0 and the gripper against Robot-Gripper 3.0.0 (blog).
- Per-engine checks: The gripper report, from validator version 2026.8.0, scores features separately for PhysX, MuJoCo, Newton and Isaac.
- Control: Grasping relies on contact and friction. The arm follows fixed waypoints with no replanning, no vision and only limited collision checking.
Read the full section
- Spec hierarchy: SimReady Foundation sets requirements in three layers. A profile (for example, robot body) contains features (for example, controllable joints), and each feature has requirements (for example, drive settings). The arm is validated against Robot-Body 4.0.0 and the gripper against Robot-Gripper 3.0.0 (blog).
- Per-engine checks: The gripper report, from validator version 2026.8.0, scores features separately for PhysX, MuJoCo, Newton and Isaac. Passing for one engine doesn't mean passing for all of them (validation JSON).
- Control: Grasping relies on contact and friction. The arm follows fixed waypoints with no replanning, no vision and only limited collision checking.
- Reproducibility: Asset commits are pinned, but the agent writes the controller code, so NVIDIA warns that code and results may differ between runs.
Claims and evidence
- The arm and gripper pass PhysX driven-joint checks. Partly confirmed.
- The library is "2,000+ assets". Coverage disagrees: Unite.AI also gives 3,599 assets across all categories. Both figures trace back to NVIDIA materials.
- Pick-and-place succeeds, with placement error checked. The blog reports no placement-error numbers, and no independent reproduction was found.
Read the full section
- The arm and gripper pass PhysX driven-joint checks. Partly confirmed. The gripper report shows PhysX-variant passes, but also 7 failures, including FET_000_PHYSX and some MuJoCo and Newton variants (validation JSON).
- The library is "2,000+ assets". Coverage disagrees: Unite.AI also gives 3,599 assets across all categories. Both figures trace back to NVIDIA materials.
- Pick-and-place succeeds, with placement error checked. This comes only from NVIDIA's own demo video and blog. The blog reports no placement-error numbers, and no independent reproduction was found.
Context and prior work
- SimReady extends NVIDIA's OpenUSD and Omniverse asset work. Approved assets are pinned to Hugging Face commits, and reviews go through the Apache-2.0 simready-central GitHub repo.
- NVIDIA USD Search provides natural-language search across the library, and Hugging Face has added it to its own platform search (HF announcement).
- Agent skills are version-controlled Markdown procedures. They ship in Isaac Sim's
skills/directory for source, pip and binary installs (Agent Skills docs).
Read the full section
- SimReady extends NVIDIA's OpenUSD and Omniverse asset work. Approved assets are pinned to Hugging Face commits, and reviews go through the Apache-2.0 simready-central GitHub repo.
- NVIDIA USD Search provides natural-language search across the library, and Hugging Face has added it to its own platform search (HF announcement).
- Agent skills are version-controlled Markdown procedures. They ship in Isaac Sim's
skills/directory for source, pip and binary installs (Agent Skills docs).
Limitations, safety and contested findings
- NVIDIA lists these limits itself (blog):
- the can always starts in a known position, and the robot is anchored
- the can doesn't deform and has no liquid inside, so crushing isn't tested
Read the full section
- NVIDIA lists these limits itself (blog):
- the can always starts in a known position, and the robot is anchored
- the can doesn't deform and has no liquid inside, so crushing isn't tested
- the grip-effort cap is not recommended by the manufacturer
- the results don't show robustness to uncertain object positions, different payloads, or real hardware
- Security: NVIDIA's docs advise running skill-driven agents only in environments you trust, with the Python server bound to 127.0.0.1 (Agent Skills docs).
- The asset-count figures conflict, and one gripper link in the blog points to the wrong model, as noted above.
Business and practitioner implications
- Equipment makers (OEMs): Publishing validated digital twins on Hugging Face makes their hardware easier to find and test in simulation (HF announcement).
- Integrators: Asset prep work gets smaller, but tuning the task (grip effort, waypoints, contact) is still your job.
- Teams using agents: The workflow depends on a coding agent sending Python to a live simulator.
Read the full section
- Equipment makers (OEMs): Publishing validated digital twins on Hugging Face makes their hardware easier to find and test in simulation (HF announcement).
- Integrators: Asset prep work gets smaller, but tuning the task (grip effort, waypoints, contact) is still your job. Read each asset's validation report for your physics engine before using it.
- Teams using agents: The workflow depends on a coding agent sending Python to a live simulator. Review the generated code, sandbox the agent, and log your pinned asset commits so runs can be compared.