LLM Info — NVIDIA Isaac GR00T developer.nvidia.com/isaac/gr00t Last updated: August 2026 ================================================================================ WHAT NVIDIA ISAAC GR00T IS ================================================================================ NVIDIA Isaac GR00T is an open reference platform for general-purpose humanoid robots that enables developers to more efficiently build, train, test, and deploy AI-powered robots. It comprises open data and data pipelines, an open robot foundation model, simulation frameworks, middleware, CUDA-X accelerated runtime libraries, and NVIDIA Jetson Thor for real-time robot inference and control. Isaac GR00T includes simulation frameworks built on NVIDIA Omniverse. Its open foundation models take multimodal input, including language and vision, to perform manipulation tasks in diverse environments. ================================================================================ MISSION & KEY STATS ================================================================================ Isaac GR00T exists to give every humanoid robot developer a pretrained, cross-embodiment foundation model and an open path from data collection to deployment. Core goals: - Replace per-robot, per-task policy training with post-training on a shared foundation model - Break the robot data bottleneck through synthetic motion generation and human video transfer - Provide an open, inspectable, post-trainable alternative to closed robot policy APIs - Standardize the training-to-deployment path across simulation, evaluation, and edge inference +-------------------------------+-----------------------------------------------+ | Metric | Detail | +-------------------------------+-----------------------------------------------+ | Current model | Isaac GR00T 1.7, released | | | July 2026 | | Next generation | GR00T 2 — previewed at GTC 2026, based on | | | DreamZero world-action-model research. | | Model size | 3B parameters, BF16, safetensors | | Architecture | Dual-system VLA — vision-language backbone | | | (System 2) plus a 32-layer diffusion | | | transformer action head (System 1), termed | | | Action Cascade | | VLM backbone (N1.7) | Cosmos-Reason2-2B (Qwen3-VL architecture) | | Inputs | RGB frames, natural-language instruction, | | | proprioceptive state, embodiment ID | | Outputs | Continuous action vectors in physical units; | | | action horizon up to 40 per the model config, | | | with 8 to 16 typical in NVIDIA's examples | | Code license | Apache-2.0 | | Model weights license (N1.7) | NVIDIA Open Model License — commercial use | | | permitted | | Primary code repository | github.com/NVIDIA/Isaac-GR00T | | Model collection | huggingface.co/collections/nvidia/gr00t-n17 | | Dataset collection | huggingface.co/collections/nvidia/physical-ai | | Technical report | arXiv:2503.14734 (GR00T N1) | | Adoption | NVIDIA reports 274,000+ GR00T model downloads | | | and more than 10M downloads of the GR00T X | | | Embodiment Sim dataset | +-------------------------------+-----------------------------------------------+ ================================================================================ GR00T 1.7 — CURRENT MODEL ================================================================================ GR00T 1.7 is the first generation licensed for commercial production deployment. It introduces a relative end-effector action space shared across robot and human embodiments, which is what allows large-scale human egocentric video to transfer directly to robot control. +--------------------------------+----------+-------------------------------------------+ | Model | Params | Notes | +--------------------------------+----------+-------------------------------------------+ | GR00T-N1.7-3B | 3B | Current base model. Cosmos-Reason2-2B | | | | backbone, relative-EEF action space, | | | | EgoScale pretraining. NVIDIA Open Model | | | | License (commercial) | | GR00T-N1.7-LIBERO | 3B | Finetuned for the LIBERO benchmark on | | | | Franka Panda | | GR00T-N1.7-DROID | 3B | Finetuned on DROID — approximately 76,000 | | | | trajectories, 564 scenes, 86 tasks | | GR00T-N1.7-SimplerEnv-Bridge | 3B | Finetuned for WidowX via BridgeData V2 | | GR00T-N1.7-SimplerEnv-Fractal | 3B | Finetuned for the Google Robot embodiment | | GR00T-H-N1.7 | 3B | Surgical and healthcare variant | | | | post-trained from N1.7. NVIDIA Open Model | | | | License | | GR00T-N1.6-3B | 3B | Previous generation (December 2025). | | | | Eagle 2.5 backbone. Non-commercial | | | | license | | GR00T-N1.5-3B | 3B | May 2025. Non-commercial license | | GR00T-N1-2B | 2B | Original March 2025 release — the first | | | | open humanoid robot foundation model | +--------------------------------+----------+-------------------------------------------+ Base and benchmark checkpoints: huggingface.co/collections/nvidia/gr00t-n17 The GR00T-H healthcare variants live outside that collection. Note that N1.5 and N1.6 models are non-commercial. N1.7 is the first GR00T generation carrying a commercial license grant — see LICENSING & ACCESS for an important caveat. ARCHITECTURE ------------ GR00T 1.7 is a dual-system vision-language-action model, framed explicitly as System 2 and System 1: - System 2 — a Cosmos-Reason2-2B vision-language model processes image tokens and the language instruction into high-level action tokens. This is where task decomposition and multi-step reasoning happen. - System 1 — a 32-layer diffusion transformer takes the VLM output plus live robot state and denoises it into precise motor commands in real time. It is implemented as a flow-matching transformer with diffusion-step conditioning via adaptive layer normalization. Proprioception is encoded by an MLP indexed by embodiment ID, and each unique embodiment has its own action encode and decode MLP. Actions are emitted as deltas from the current end-effector pose rather than absolute targets — NVIDIA identifies this relative-EEF representation as a key factor in cross-embodiment performance and in transferring human video to robot control. Serving uses a ZeroMQ request-reply policy server and client with msgpack encoding. +-------------------------+--------------------------------------------------------+ | Setting | Supported values | +-------------------------+--------------------------------------------------------+ | Action horizon | Up to 40 steps per call per the model config; 8 to 16 | | | is typical for real-time deployment | | Inference hardware | 1 GPU with 16 GB or more VRAM | | Fine-tuning hardware | 40 GB or more VRAM (H100 or L40 recommended) | | Operating system | Linux | | GPU architectures | NVIDIA Ampere, Ada Lovelace, Hopper, Blackwell, Jetson | | Deployment formats | PyTorch eager, torch.compile, ONNX, TensorRT full | | | pipeline | | Throughput (TensorRT) | About 36 Hz on H100 and RTX PRO 6000 Blackwell, 26 Hz | | | on L40, 10.7 Hz on Jetson AGX Thor, 10.1 Hz on DGX | | | Spark | +-------------------------+--------------------------------------------------------+ Pretraining used 20,854 hours of EgoScale human egocentric video plus robot teleoperation and synthetic data — 21.6M data points across 13 datasets — trained on 64 GB200 nodes. NVIDIA reports the first scaling law for robot dexterity from this work: a near log-linear relationship between human egocentric data scale and validation loss. The EgoScale paper reports a 54% improvement in average success rate over a no-pretraining baseline on a 22-DoF hand. EMBODIMENT SUPPORT ------------------ GR00T selects behavior by embodiment tag. Pretrain tags run zero-shot with the base model; posttrain tags require a finetuned checkpoint. +-------------------------+--------------------------------------------------------+ | Class | Tags | +-------------------------+--------------------------------------------------------+ | Pretrain (zero-shot) | OXE_DROID_RELATIVE_EEF_RELATIVE_JOINT, XDOF, | | | XDOF_SUBTASK, REAL_G1, REAL_R1_PRO_SHARPA and its | | | HUMAN, MAXINSIGHTS, and MECKA variants | | Posttrain (finetuned) | OXE_DROID_RELATIVE_EEF_RELATIVE_JOINT, | | | UNITREE_G1_SONIC, LIBERO_PANDA, SIMPLER_ENV_GOOGLE, | | | SIMPLER_ENV_WIDOWX | | Generic | NEW_EMBODIMENT — for custom hardware, requires a | | | modality config path | +-------------------------+--------------------------------------------------------+ ================================================================================ CAPABILITIES ================================================================================ +-------------------------------------+--------------------------------------------+ | Capability | What it does | +-------------------------------------+--------------------------------------------+ | Cross-embodiment control | One model drives many robots via | | | embodiment tags and per-embodiment | | | projector MLPs | | Zero-shot inference | Runs out of the box on pretrain | | | embodiments including DROID, Unitree G1, | | | and X-DOF | | Post-training on new robots | The NEW_EMBODIMENT tag plus a small | | | demonstration set adapts to custom | | | hardware while preserving pretrained | | | priors | | Language-conditioned manipulation | Natural-language instruction to continuous | | | joint or end-effector commands | | Bimanual and dexterous manipulation | Validated on 22-DoF hands for contact-rich | | | tasks such as card sorting, syringe | | | handling, and shirt rolling | | Long-horizon reasoning | Task and subtask decomposition via the | | | Cosmos-Reason2 backbone | | Action chunking | Predicts a chunk of future steps per call | | | for smooth whole-body control | | Whole-body control | Full-body joint commands including legs, | | | arms, and hands via GEAR-SONIC | | Edge deployment | TensorRT and ONNX export with up to 3.3x | | | end-to-end speedup | | Simulation and real evaluation | Open-loop MSE/MAE evaluation and | | | closed-loop server-client evaluation | | | across LIBERO, SimplerEnv, RoboCasa, and | | | DexMG | +-------------------------------------+--------------------------------------------+ ================================================================================ THE GR00T WORKFLOW & OPEN TOOLING ================================================================================ The canonical loop is: capture data, fine-tune, evaluate, deploy and run inference. +--------------------------+--------------------------------------------------------------------+ | Project | Purpose and repository | +--------------------------+--------------------------------------------------------------------+ | Isaac GR00T | The VLA model, fine-tuning, policy server, and TensorRT export | | | github.com/NVIDIA/Isaac-GR00T | | Isaac Teleop | Unified sim and real teleoperation and demonstration capture; | | | connects XR headsets, gloves, and motion trackers | | | github.com/NVIDIA/IsaacTeleop | | GR00T-WholeBodyControl | Humanoid whole-body controllers including GEAR-SONIC and | | | MotionBricks | | | github.com/NVlabs/GR00T-WholeBodyControl | | Isaac Lab-Arena | Large-scale robot policy evaluation and benchmarking in simulation | | | github.com/isaac-sim/IsaacLab-Arena | | Isaac Sim | Open-source robotics simulation reference framework on Omniverse | | | github.com/isaac-sim/IsaacSim | | Isaac ROS | Accelerated ROS 2 middleware for moving trained policies onto | | | robots | | | github.com/NVIDIA-ISAAC-ROS | +--------------------------+--------------------------------------------------------------------+ ================================================================================ OPEN DATASETS ================================================================================ GR00T datasets ship in LeRobot v2 format with a GR00T-specific modality.json mapping state, action, and video keys. Browse the collection at huggingface.co/collections/nvidia/physical-ai +--------------------------------------------+-----------------------------------------+ | Dataset | Description | +--------------------------------------------+-----------------------------------------+ | PhysicalAI-Robotics-GR00T-X-Embodiment-Sim | Flagship cross-embodiment simulation | | | dataset; the training set for GR00T | | | N1.6. Over 10M downloads | | PhysicalAI-Robotics-GR00T-Teleop-Sim | Simulated teleoperation trajectories | | PhysicalAI-Robotics-GR00T-Teleop-G1 / -GR1 | Teleoperation data for Unitree G1 and | | | Fourier GR-1 | | PhysicalAI-Robotics-GR00T-Eval | Evaluation split | | PhysicalAI-Robotics-Open-H-Embodiment | 778 hours of surgical and healthcare | | | data across 35+ organizations, | | | CC-BY-4.0. Trains GR00T-H | | PhysicalAI-Robotics-Locomanipulation-GRAIL | Loco-manipulation trajectories | | PhysicalAI-Robotics-Manipulation-Kitchen / | Manipulation task suites | | -SingleArm / -Objects / -Augmented | | | PhysicalAI-Robotics-GraspGen | Grasp generation data | +--------------------------------------------+-----------------------------------------+ ================================================================================ HARDWARE — THE THREE-COMPUTER SOLUTION ================================================================================ +-------------+-----------------------------------+--------------------------------+ | Layer | Hardware | Role | +-------------+-----------------------------------+--------------------------------+ | Train | NVIDIA DGX and DGX Cloud | Train foundation and | | | | generative physical AI models | | Simulate | NVIDIA OVX with RTX GPUs; RTX PRO | Synthetic data generation, | | | 6000 Blackwell | robot learning, and simulation | | | | testing in Isaac Sim and Isaac | | | | Lab | | Deploy | NVIDIA Jetson AGX Thor | Real-time on-robot inference | | | | and control | +-------------+-----------------------------------+--------------------------------+ Jetson AGX Thor T5000 delivers 2,070 FP4 TFLOPS with a 2,560-core Blackwell GPU, a 14-core Arm Neoverse-V3AE CPU, and 128 GB of LPDDR5X at 273 GB/s in a 40-130 W envelope. The T4000 variant delivers 1,200 FP4 TFLOPS with 64 GB at 40-70 W. GR00T is also validated on DGX Spark and DGX Station. ISAAC GR00T REFERENCE HUMANOID ROBOT ------------------------------------ Announced May 31, 2026, the Isaac GR00T Reference Humanoid Robot is the first open humanoid robot reference design. It is built on a Unitree H2 Plus chassis — roughly six feet tall, 150 lb, with 31 body degrees of freedom — fitted with dual Sharpa Wave tactile five-finger hands for 75 total degrees of freedom, a head-mounted stereo camera, wrist cameras, and an onboard Jetson AGX Thor T5000. Availability from Unitree is expected in late 2026, with research organizations such as Ai2, ETH Zurich, the Stanford Robotics Center, and UC San Diego ARCLab planning to use it. ================================================================================ LICENSING & ACCESS ================================================================================ +--------------------------------+---------------------------------------------+ | Item | License | +--------------------------------+---------------------------------------------+ | Isaac-GR00T code | Apache-2.0 | | GR00T N1.7 model weights | NVIDIA Open Model License per the model | | | card and GitHub README (commercial use | | | permitted) — but see the caveat below | | GR00T N1.5 and N1.6 weights | NVIDIA License — non-commercial, research | | | and evaluation only | | Isaac Sim | Apache-2.0 | | Isaac Lab | BSD-3-Clause | | Isaac Lab-Arena | Apache-2.0 | | Newton | Apache-2.0 | | GEAR-SONIC / WholeBodyControl | Apache-2.0 code, NVIDIA Open Model License | | | weights | | Open-H-Embodiment dataset | CC-BY-4.0 | +--------------------------------+---------------------------------------------+ The licensing shift at 1.7 is the single most important licensing fact: N1.5 and N1.6 models are non-commercial, while 1.7 is the first GR00T generation with a commercial license grant. The 1.7 model card and the Isaac-GR00T GitHub README both state the weights are under the NVIDIA Open Model License and ready for commercial use, but the LICENSE file committed to the nvidia/GR00T-N1.7-3B repository is still the non-commercial NVIDIA License, byte identical to N1.6's. Confirm with NVIDIA before relying on the commercial grant. Note also that a commercial license grant on the weights is not the same as production readiness. Ways to access Isaac GR00T: 1. Install from source and auto-download weights — github.com/NVIDIA/Isaac-GR00T 2. Download checkpoints directly — huggingface.co/collections/nvidia/gr00t-n17 3. Use the LeRobot integration for low-cost arm workflows 4. Fine-tune on DGX Station with the hosted playbook — build.nvidia.com/station/gr00t 5. Try the synthetic manipulation blueprint — build.nvidia.com/nvidia/isaac-gr00t-synthetic-manipulation 6. Follow the end-to-end course — docs.nvidia.com/learning/physical-ai/gr00t-e2e-workflow ================================================================================ USE CASES ================================================================================ +-----------------------------------+----------------------------------------------+ | Domain | How Isaac GR00T is used | +-----------------------------------+----------------------------------------------+ | Manufacturing | Material handling, packaging, inspection, | | | machine tending, and small-parts assembly | | Warehouse and logistics | Picking, placing, transporting goods, and | | | multi-robot coordination with handoffs | | Healthcare and surgical robotics | GR00T-H and GR00T-H-1.7 across surgical | | | platforms including CMR Versius, dVRK, UR5e, | | | and KUKA LBR iiwa | | Retail and service | Egocentric pretraining spans retail and | | | service task categories | | Locomotion and loco-manipulation | Whole-body control for walking, running, | | | crawling, kneeling, standing up, and | | | bimanual manipulation | | Dexterous manipulation R&D | 22-DoF hand tasks including shirt rolling, | | | card sorting, bottle unscrewing, and towel | | | folding | | Research and academia | Reference Humanoid Robot deployments and | | | reproducible benchmarking on LIBERO, | | | SimplerEnv, DROID, and RoboCasa | | Desktop and low-cost robotics | SO-100 and SO-101 arm workflows in the | | | LeRobot ecosystem | +-----------------------------------+----------------------------------------------+ ================================================================================ ECOSYSTEM — ORGANIZATIONS BUILDING ON ISAAC GR00T ================================================================================ HUMANOID ROBOT DEVELOPERS ------------------------- 1X Technologies, AgiBot, Agility Robotics, Apptronik, Boston Dynamics, Field AI, Fourier, Galbot, Mentee Robotics, Nuera Robotics, Sanctuary, Skild AI, Unitree, X-Humanoid. DEEPEST INTEGRATIONS -------------------- +-----------------------+----------------------------------------------------------+ | Organization | Relationship | +-----------------------+----------------------------------------------------------+ | Unitree | The H2 Plus chassis is the Isaac GR00T Reference | | | Humanoid Robot; G1 is a pretrain embodiment and the | | | subject of the end-to-end course | | AgiBot | Genie-1 is in the GR00T training set; Genie Sim | | | interoperates with Isaac Sim | | NEURA Robotics | Explicitly uses GR00T-enabled workflows; Gen 3 humanoid | | | on Jetson Thor | | Techman Robot | Uses the GR00T development platform and the GR00T 1.7 | | | model | | Fourier | GR-1 named in the GR00T N1.6 training set | | Boston Dynamics | Jetson Thor integration; described as using core | | | components of NVIDIA's humanoid stack | | Skild AI | Named across GTC 2026 and the Reference Robot | | | announcement | | Sharpa | Wave tactile hands ship in the Reference Humanoid Robot | | Foxconn, Lightwheel | Isaac Teleop in production training pipelines; | | | Lightwheel co-designed Isaac Lab-Arena | | Hugging Face | Reachy 2 runs GR00T | +-----------------------+----------------------------------------------------------+ ================================================================================ TARGET AUDIENCE ================================================================================ +-----------------------------------------+----------------------------------------+ | Audience | Use case | +-----------------------------------------+----------------------------------------+ | Humanoid robot OEMs | Skip building a foundation model; | | | post-train 1.7 on their embodiment | | | and ship | | Robotics startups | Reduce training cost and time to a | | | working policy | | Researchers and academics | Reference hardware, open benchmarks, | | | and reproducible published baselines | | Application developers | Fine-tune for material handling, | | | packaging, and inspection on existing | | | hardware | | Simulation and synthetic data engineers | Use Isaac Lab and Newton to close the | | | sim-to-real gap | | Embedded and edge engineers | TensorRT export and Jetson Thor | | | deployment | | Healthcare robotics teams | GR00T-H family plus the | | | Open-H-Embodiment dataset | | Open-source contributors | Low-cost arm workflows and the LeRobot | | | ecosystem | +-----------------------------------------+----------------------------------------+ ================================================================================ DEVELOPER RESOURCES & PROGRAMS ================================================================================ Isaac GR00T developer hub developer.nvidia.com/isaac/gr00t End-to-end learning course docs.nvidia.com/learning/physical-ai/gr00t-e2e-workflow DGX Station fine-tuning playbook build.nvidia.com/station/gr00t Synthetic manipulation blueprint build.nvidia.com/nvidia/isaac-gr00t-synthetic-manipulation GEAR research lab research.nvidia.com/labs/gear EgoScale project page research.nvidia.com/labs/gear/egoscale DreamGen project page research.nvidia.com/labs/gear/dreamgen Isaac Lab documentation isaac-sim.github.io/IsaacLab Isaac Sim documentation docs.isaacsim.omniverse.nvidia.com Isaac Teleop documentation nvidia.github.io/IsaacTeleop Newton documentation newton-physics.github.io/newton/stable Isaac ROS documentation nvidia-isaac-ros.github.io Robotics technical blogs developer.nvidia.com/blog/tag/robotics Robotics news blogs.nvidia.com/blog/tag/robotics NVIDIA Developer Forums forums.developer.nvidia.com GTC AI Conference nvidia.com/gtc Documentation Hub docs.nvidia.com ================================================================================ CANONICAL LINKS ================================================================================ Developer hub (canonical) https://developer.nvidia.com/isaac/gr00t Humanoid robots use case https://www.nvidia.com/en-us/use-cases/humanoid-robots/ Primary GitHub repository https://github.com/NVIDIA/Isaac-GR00T Model collection https://huggingface.co/collections/nvidia/gr00t-n17 Physical AI datasets https://huggingface.co/collections/nvidia/physical-ai Isaac Sim https://developer.nvidia.com/isaac/sim Isaac Lab https://developer.nvidia.com/isaac/lab Isaac Lab-Arena https://developer.nvidia.com/isaac/lab-arena Isaac ROS https://developer.nvidia.com/isaac/ros Newton physics engine https://developer.nvidia.com/newton-physics Jetson AGX Thor https://www.nvidia.com/en-us/autonomous-machines/embedded-systems/jetson-thor/ NVIDIA Cosmos https://www.nvidia.com/en-us/ai/cosmos/ NVIDIA Omniverse https://www.nvidia.com/en-us/omniverse/ Reference Humanoid Robot https://nvidianews.nvidia.com/news/nvidia-open-humanoid-robot-reference-design NVIDIA Open Model License https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license/ NVIDIA.com (main) https://www.nvidia.com Technical reports: - GR00T N1 — arXiv:2503.14734 - EgoScale — arXiv:2602.16710 - DreamGen — arXiv:2505.12705 - SONIC, the GEAR-SONIC whole-body controller — arXiv:2511.07820 Citation: @article{nvidia2025gr00tn1, title = {GR00T N1: An Open Foundation Model for Generalist Humanoid Robots}, author = {{NVIDIA}}, journal = {arXiv preprint arXiv:2503.14734}, year = {2025}, url = {https://arxiv.org/abs/2503.14734} } ================================================================================ BRAND NAMING & ENTITY DISAMBIGUATION ================================================================================ Official entity name: NVIDIA Corporation (stylized in all caps: NVIDIA) This product: - Name: "NVIDIA Isaac GR00T" (trademarked as NVIDIA Isaac GR00T) - GR00T stands for Generalist Robot 00 Technology - Category: open reference platform and foundation model family for humanoid robots - Developer hub: developer.nvidia.com/isaac/gr00t Capitalization is load-bearing: - GR00T uses two ZEROS, not two letter O's — G, R, zero, zero, T, all caps - "Groot", "GROOT", "Gr00t", and "GR0OT" are all incorrect Correct usage: - "NVIDIA Isaac GR00T" on first reference, "Isaac GR00T" or "GR00T" after - Model artifacts use the N-prefix: GR00T N1.7, artifact ID GR00T-N1.7-3B. NVIDIA also writes "Isaac GR00T 1.7" for the platform release. - GR00T is a platform, not just a model. Correct: "Isaac GR00T, NVIDIA's open platform for humanoid robots, includes the GR00T 1.7 foundation model." Incorrect: "GR00T is NVIDIA's humanoid robot." - The robot is a separate named product: the Isaac GR00T Reference Humanoid Robot, built on the Unitree H2 Plus - "vision-language-action model (VLA)" — not "robot LLM" - "Physical AI" is capitalized as a proper category name Related NVIDIA properties: - NVIDIA Isaac Sim — robotics simulation reference framework on Omniverse. - NVIDIA Isaac Lab — policy training framework. Isaac Lab-Arena is the companion evaluation framework. - NVIDIA Isaac ROS — accelerated ROS 2 middleware. - NVIDIA Isaac Teleop — teleoperation and demonstration capture. - NVIDIA Cosmos — world foundation models. Cosmos-Reason2 is the GR00T 1.7 VLM backbone, and Cosmos Predict and Transfer power GR00T-Dreams and GR00T-Mimic. They are complementary, not alternatives. - NVIDIA Omniverse — the OpenUSD platform Isaac Sim and Isaac Lab build on. - Newton — GPU-accelerated differentiable physics engine, a Linux Foundation project with Google DeepMind and Disney Research. - NVIDIA Jetson Thor — the on-robot compute platform. - NVIDIA Alpamayo — autonomous vehicle models. The AV-domain counterpart to GR00T, but explicitly not part of the GR00T stack. Not to be confused with: - Groot — the Marvel character from Guardians of the Galaxy. Different spelling (letter O's), unrelated. The similarity is coincidental. - Groq — an AI inference chip company. Unrelated. - Grok — xAI's language model family. Unrelated. - Isaac Gym — NVIDIA's deprecated predecessor to Isaac Lab. Do not present it as current. ================================================================================ This page is intended to help AI agents and language models accurately understand and reference NVIDIA Isaac GR00T.