GPU-accelerated simulation, training and inference.
Ipnops provides its own I/O for robots, machines, infrastructure and software, and runs GPU-accelerated simulation, training and inference as jobs. Scenes, policies, datasets and readings move in open formats: OpenUSD, ONNX and MCAP. Ipnops is part of Living Worlds Lab, an NVIDIA Inception member.
From a training task to a twin.
A task runs in a runtime with GPU access and no network. Its outputs reach the twin as OpenUSD stages and ONNX policies, and the command path to a machine stays behind the gate.
What runs, by area.
Capabilities, not product lists. Some components run under development, evaluation or research licences and do not back a paid deliverable until licensed; the maker's licence for that use is put in place first.
Scenes, twins and assets exchange as OpenUSD stages with physics and semantic labels, and simulation-ready assets are imported behind a licence gate. See below.
GPU physics for cloth, rigid and deformable bodies, run as jobs in rootless containers with GPU access.
Training tasks train, play, roll out and export policies to ONNX, and fine-tune policy models on open robot datasets.
Video and world models for generation and transfer, served in runtimes behind the maker's guardrail.
Motion plans are written as MCAP trajectories for review before anything is issued; perception reads images and sensor streams.
Weather forecasts and radio maps from public inputs, as served jobs.
Room impulse responses and a GPU wave-equation engine.
Workers advertise the GPUs they hold; jobs name what they need (GPU class, memory, cores, operating system) and place on them.
A learned-policy runtime contract that never actuates, simulation backend choice, task descriptors, surrogate-model hosting, typed world-model jobs and a detection schema for sensor pipelines. Built and tested in shadow mode.
Assets with their licence attached.
Browse a catalogue of simulation-ready OpenUSD assets, read each asset's licence, import an asset after you accept it, and get it back checked, with its licence and provenance.
An index of simulation-ready OpenUSD assets, refreshed hourly. Search by keyword and dataset.
Nothing downloads until you name the asset's licence exactly. An unknown or other licence allows evaluation under the maker's terms only.
Every file is hash-checked against the maker's bill of materials. Paths, file types, sizes and time are capped. USD is read in a sandboxed process with no network and no program execution, and an import refuses if the sandbox cannot be applied.
Ipnops checks layout, units, default prim, masses and references against the public specification, recomputes mass from the rigid bodies and keeps it beside the maker's declared value.
The kept package carries its licence flags, the accepted licence and time, the pinned commit and hashes, and the maker's own validation results.
Four tools over MCP: catalogue, describe, import and validate. See Developers.
Tasks and models go out to a runtime; USD stages, ONNX policies, trajectories and videos come back as signed outputs.
A job names a kind. No code travels with it: each machine runs only the kinds its owner listed. The job runs in a rootless container with GPU device access and no network.
Outputs are inputs to a twin, not commands. A policy reaches a machine only through the interface: forecast in the twin, purpose check, command gate, approval.
Run it on your GPUs or ours.
Jobs run on the Ipnops fleet or on machines you bring.