IllustrationNo official image on file. Showing a schematic category illustration.
Open reasoning vision-language-action model for humanoids, released with NVIDIA's CES 2026 physical-AI model wave.
Key specs
| Product type | Open VLA with full-body control for humanoids |
|---|
Technical details
Isaac GR00T N1.6 is NVIDIA's open 3B-parameter VLA for generalized humanoid skills, combining a vision-language model with a flow-matching DiT head that denoises continuous actions. Image tokens from all camera views are concatenated with language tokens; proprioception and actions use per-embodiment MLP encoders/decoders, so one model handles varying views and DoF. N1.6 modifies the VLM-to-DiT connector, trains jointly with flow-matching and world-modeling objectives, and per NVIDIA uses Cosmos Reason for better reasoning and full-body humanoid control. Weights are on Hugging Face and the model is integrated in LeRobot for fine-tuning.
| Architecture | Vision-language foundation model + flow-matching diffusion transformer (DiT) action head: SigLip2 vision transformer and T5 text encoder, embodiment-indexed MLP proprioception encoders, DiT with AdaLN interleaving self-attention over proprio/actions with cross-attention to vision-language tokens; trained jointly with flow-matching and world-modeling objectives; uses Cosmos Reason for reasoning. |
|---|---|
| Params | 3B |
| Inputs outputs | Inputs: variable number of RGB camera frames (any resolution), robot proprioception (float vector), language instruction; outputs: continuous action vectors sized to the embodiment's DoF. |
| Training data | Mixture of bimanual, semi-humanoid and an expansive humanoid dataset, combining real captured data and synthetic data generated with NVIDIA Isaac GR00T Blueprint components. |
| Embodiments supported | Cross-embodiment; humanoids (full-body control), bimanual and semi-humanoid robots; adaptable via post-training. |
| License weights | NVIDIA License (per Hugging Face model card) |
| License code | Apache-2.0 (Isaac-GR00T repository) |
| Hardware requirements | NVIDIA GPUs: Ampere, Lovelace, Hopper, Blackwell, Jetson; Linux; PyTorch runtime. |
| How to try | github.com/…/Isaac-GR00T |
| Sources |
