"Supported" is doing quiet work in that release note. In an open library a deprecated model rarely disappears, because weights stay downloadable and old commits stay checked out. What moves is the maintained path, and neither NVIDIA's post nor the Hugging Face guide gives a deprecation date, a migration procedure for policies already fine-tuned on N1.5, or a statement about whether N1.5 checkpoints still load once 1.7 takes the slot [20]. Until a team tests that on its own hardware the size of the job is unknown, which is the worst shape a cost can have in a quarter that is already allocated.
The reason a fine-tune does not simply carry across is that the policy is a function of its base. NVIDIA describes 1.7 as delivering improved performance and expanded manipulation over N1.5 [2], which is a direction of travel rather than a specification of how behaviour differs on your task. The release does promise that models trained anywhere can be deployed through LeRobot while maintaining the same benchmarked performance [13], but that is portability of a trained artifact into deployment, not equivalence between two base models. So the work is post-training again and evaluating again against your own suite, and the acceptance numbers you showed your steering group last quarter describe a model that is no longer the library default.
Here is the board-deck version. NVIDIA counts 3 million robotics developers and Hugging Face counts 16 million AI builders [14], the collaboration therefore addresses something like 19 million people [15], and a shared stack lowers the entry cost to physical AI work that NVIDIA itself calls costly and fragmented [21]. It is incomplete in the ordinary way: adding 3 to 16 double counts anyone who sits in both populations, since neither source states the overlap [15], and the figure an operator actually needs, how many teams have hardware-attached policies resting on N1.5, appears nowhere.
A skeptic would say this is a version bump in a fast-moving research library, and that nobody with robots in revenue tracks LeRobot's default model. Partly right, and the counter-evidence is in the install instructions. The walkthrough runs a real-world task on an SO-101 arm [12] with demonstrations collected through either a VR headset or the leader arm [19]; the setup pins isaacteleop to ~=1.3.131 alongside scipy [9]; DGX Spark users are routed to CUDA 13 builds of torch 2.11.0 [8]; Jetson Thor deployment through LeRobot's Reachy 2 is already part of the surrounding kit [18]. Pinned stacks with hardware attached are precisely where an upstream default change consumes a week of somebody's calendar.
The distinction worth holding is between this week and this decade. This week the choice is narrow: book the revalidation, or hold N1.5 and accept that you maintain it yourself. Over a longer horizon the question is whose release cadence sets your robot stack's cadence, and Cosmos 3 is already named as the next NVIDIA model arriving in LeRobot [5]. Teams whose durable asset is demonstration data in a standard interoperable format, which is what Isaac Teleop exists to produce [4], will pay less each time that cadence turns over than teams whose asset is one tuned checkpoint.
