NVIDIA Isaac GR00T N1.5
NVIDIA, United States · May 2025
- Parameters
- 2.7 B
- 2,724,163,520 parameters in BF16 according to the safetensors metadata of nvidia/GR00T-N1.5-3B on Hugging Face.
- GPU memory
- 5.1 GB
- weights at bf16, computed. Fits an 8 GB card
- Inference latency
- not published
- per action step
- Weights
- NVIDIA OneWay NC
- NVIDIA OneWay Noncommercial License for the weights.
What it is
GR00T N1.5 is a data and training upgrade rather than a new architecture, and it is the release where the family became usable. Freezing the Eagle 2.5 VLM instead of fine-tuning it stopped the model from forgetting language, taking real-robot language following from 46.6 to 93.3 percent. FLARE added learning from action-less human video and DreamGen added synthetic behaviours, together tripling low-data RoboCasa performance. What is missing is any measurement a deployment engineer needs: no latency figure, no VRAM table, and no comparison against anything outside NVIDIA's own lineup. The license is noncommercial, so despite being the best-documented mid-generation model here it is a research asset only.
Architecture
- Backbone
- Eagle 2.5, kept frozen during both pre-training and fine-tuning. The VLM is 2.1B parameters.
- Action head
- flow matching diffusion transformer (16 layers) with adaptive layer normalisation, trained jointly with a world-modelling objective. Adds the FLARE future-latent-representation-alignment objective alongside the flow matching loss so the model can learn from action-less human egocentric video.
- Action chunk
- 16 steps per forward pass
- Parameters
- 2,724,163,520 parameters in BF16 according to the safetensors metadata of nvidia/GR00T-N1.5-3B on Hugging Face. The repository name says 3B, which rounds up. The VLM component alone is listed as 2.1B in the GEAR blog's grounding comparison table.
- Pretraining data
- Extends the N1 data pyramid with DreamGen synthetic neural trajectories, which generate novel behaviours beyond what teleoperation covered, plus human egocentric video consumed through the FLARE objective. Public sim component: nvidia/PhysicalAI-Robotics-GR00T-X-Embodiment-Sim with 1.34M examples. NVIDIA describes the N1.x base checkpoints as pre-trained on 10k+ hours of robot data.
- Embodiments
- Fourier GR-1 humanoid, Unitree G1, single-arm robots with end-effector control via the OXE_DROID embodiment head, humanoids with grippers via the AGIBOT_GENIE1 head, SO-100 and other arms through the NEW_EMBODIMENT path
What hardware it needs
This is the section most people came for, so it is worth being precise about which numbers are measured and which are arithmetic.
| Precision | Weights | Note |
|---|---|---|
| fp32 | 10 GB | Training master weights, rarely used for inference |
| bf16 or fp16 | 5.1 GB | The usual way these checkpoints are served |
| int8 | 2.5 GB | Quantised, expect some loss of precision on fine motions |
| int4 | 1.3 GB | Aggressive quantisation, verify success rate before trusting it |
Smallest real card that fits the bf16 weights plus a 40 percent runtime allowance: 8 GB. That is an estimate for inference, not for fine tuning. Full fine tuning also holds optimiser state and gradients, which typically costs several times the weights unless you use LoRA or a comparable adapter.
No VRAM table published for N1.5. The N1.5 README states fine-tuning and inference were tested on H100, L40, RTX 4090 and A6000, and additionally verified on an RTX 3090 Ti with CUDA 11.8. Inference was tested on RTX 3090, RTX 3090 Ti, RTX 4090 and A6000.
The authors publish no memory or latency figure for this model. Everything above is computed from the parameter count.
Reported results
Grouped by what the evaluation actually asked. Values from different suites are not comparable with each other, so each one keeps its suite and split.
Simulation
Reproducible benchmark suites. Everyone runs the same episodes, so the numbers can be compared inside a suite.
- RoboCasa Post-training in the data-limited regime, 30 demonstrations per task47.5%success rateFirst-party. GR00T N1 scored 17.4 on the identical low-data split. Note this is a different split from the 100-demos-per-task setting where N1 scored 32.1 percent, so the two are not directly comparable.source
Real world
Rollouts on physical hardware. The setups differ, so read these as evidence, not as a ranking.
- Real GR-1 humanoid manipulation, language following Real robot, language command following rate93.3%language following rateFirst-party. GR00T N1 scored 46.6 percent on the same evaluation. This is the headline improvement of the release and comes from freezing the VLM.source
- Real GR-1 humanoid manipulation Real robot, overall task success83%
- VLM grounding evaluation GR-1 grounding tasks, intersection over union40.4 IoUIoUFirst-party. The 2.1B N1.5 VLM versus Qwen2.5VL 3B at 35.5 IoU. On RefCOCOg-val the same comparison is 89.6 versus 85.2.source
- DreamGen novel-verb generalisation 12 verbs unseen in teleoperation data38.3%success rateFirst-party. GR00T N1 scored 13.1 percent. Absolute performance on genuinely novel verbs remains low.source
Fine tuned tasks
What the model reaches after being adapted to a new task or a new robot, usually from a small number of demonstrations. This is the number that matters if you bring your own data.
- Unitree G1 post-training Place one of two fruits or objects, 1000 demonstrations98.8%success rateFirst-party. GR00T N1 with the same 1000 demonstrations reached 44.0 percent. With five novel objects N1.5 dropped to 84.2 percent.source
On real hardware
Evaluated on the Fourier GR-1 humanoid for language following and overall manipulation success, and post-trained on the Unitree G1 for pick-and-place, reaching 98.8 percent on placing one of two objects with 1000 demonstrations and 84.2 percent when the objects were novel. NVIDIA also published a Jetson deployment guide for N1.5 in the repository.
Fine tuning it yourself
Fine-tune from the n1d5 branch of Isaac-GR00T using GR00T LeRobot format data. Three pre-trained embodiment heads are provided (joint-space humanoid, OXE_DROID end-effector single arm, AGIBOT_GENIE1 gripper humanoid), and custom robots go through the NEW_EMBODIMENT path. The published data-efficiency results use 30, 100 and 300 demonstrations per task, and the Unitree G1 result used 1000 demonstrations. The VLM stays frozen by default, which is what keeps fine-tuning affordable.
This is one of the models the AY-Robots training pool can fine tune for you on a dataset recorded with your own arm. The platform page carries the defaults, GPU tier and inference latency measured in that pool.
Platform page for NVIDIA Isaac GR00T N1.5Where it helps, where it does not
Strengths
- Freezing the VLM roughly doubled language following on the real GR-1, from 46.6 to 93.3 percent, and lifted overall real-robot success from 43.3 to 83.0 percent. That is the largest single-step gain anywhere in the N1.x line.
- Far better in the low-data regime. 47.5 versus 17.4 percent on RoboCasa with only 30 demonstrations per task, which is the regime most teams actually operate in.
- FLARE lets the model learn from action-less human egocentric video, and DreamGen supplies synthetic trajectories for behaviours never teleoperated.
- Verified to run on modest hardware. NVIDIA reports successful fine-tuning and inference on an RTX 3090 Ti and inference on an RTX 3090.
- Added embodiment heads for single-arm end-effector control (OXE_DROID) and gripper humanoids (AGIBOT_GENIE1), broadening it beyond joint-space humanoids.
- Still the version integrated into LeRobot's documented policy set, which lowers the integration cost for existing LeRobot users.
Limits
- Noncommercial license. Same NVIDIA OneWay Noncommercial terms as N1, so no product deployment.
- NVIDIA published no latency measurement and no VRAM table for N1.5. Hardware planning rests on a list of GPUs that were tested, without numbers.
- Novel-verb generalisation is still weak in absolute terms at 38.3 percent, and zero-shot novel-object handling was only 15.0 percent before FLARE post-training.
- All published comparisons are against GR00T N1, never against π0, π0.5 or any non-NVIDIA model, so the numbers do not place N1.5 in the wider field.
- The benchmark splits changed between N1 and N1.5, for instance 30 versus 100 demonstrations per task on RoboCasa, so cross-version reading requires care.
- Two revisions behind, and NVIDIA moved the repository to branch n1d5 for archival.
Sources
Everything on this page was taken from these documents. Where they disagree with what you read here, they win.
- https://research.nvidia.com/labs/gear/gr00t-n1_5/
- https://huggingface.co/nvidia/GR00T-N1.5-3B
- https://github.com/NVIDIA/Isaac-GR00T/tree/n1d5
- https://arxiv.org/abs/2503.14734
- https://huggingface.co/api/models/nvidia/GR00T-N1.5-3B
Entry last checked 2026-08-11.
Try a policy on a real arm
A physical SO-100 is online and free to drive in the browser, and five of the models in this arena can be fine tuned on a dataset you record with your own.