NVIDIA Isaac GR00T N1.7

NVIDIA Isaac GR00T N1.7 is the current version of NVIDIA's open vision-language-action model for robots, and the point at which the line stopped being purely a research release. Earlier versions shipped under a non-commercial NVIDIA licence. N1.7 was announced in early access on 17 April 2026 with, in NVIDIA's words, "a license that supports production use cases", and the Isaac-GR00T repository now describes it as generally available, with support and stability guarantees. The architecture changed at both ends. The vision-language backbone is Cosmos-Reason2-2B, built on the Qwen3-VL architecture, replacing the Eagle backbone used in N1.6 — NVIDIA's own reasoning model now sits inside NVIDIA's own robot policy. The action head remains a flow-matching diffusion transformer, but its depth moved in the opposite direction to the previous release: N1.6 had doubled it from 16 to 32 layers, and N1.7 halves it back to 16. Two capabilities are new — structured reasoning at task and subtask level, and finger-level actuation for contact-rich work such as small-parts assembly. The model has 3 billion parameters. Pretraining used 21.6 million data points drawn from 13 datasets, collected as a hybrid of human, robot and simulated sources, with RGB video, proprioception and actuator readings. NVIDIA also publishes the cost of that training run: 41,288 kWh and 16.949 tonnes of CO2 equivalent. The deployment figures are unusually concrete for a robot policy. With TensorRT on an H100 the full pipeline completes a step in 27.9 ms — 35.9 Hz, a 3.08x speed-up. On a Jetson Orin only the diffusion transformer can be compiled, because TensorRT 10.3 does not support the backbone engine, and the model runs at 216.5 ms per step, or 4.6 Hz. That gap is the practical cost of moving the policy onto the robot itself. Post-trained checkpoints ship alongside the base model: DROID (roughly 76,000 demonstration trajectories, about 350 hours, 564 scenes across 52 buildings, 86 manipulation tasks), LIBERO (130 language-conditioned tasks) and two SimplerEnv variants. Older versions are not withdrawn — N1.6 and N1.5 remain available in the n1d6 and n1d5 branches, and both keep their own entries in this catalogue where they exist. One inconsistency is worth recording, because both statements come from the same company: NVIDIA's research page for N1.6 described that version as using an internal Cosmos-2B variant, while the N1.7 repository states that N1.6 used the Eagle backbone.

▸News

▸Videos

No videos yet.

▸Reviews