Open source

Reproducible by default.

We release the benchmarks, datasets, and software behind our papers, so that the results can be checked and built on by others. Each release below is tied to a publication where one exists.

cerebrum-core Software

The reference implementation of our perceive–plan–act stack: a shared representation layer with perception, planning, and control back ends that read and write the same belief state.

PythonPyTorchLicense: Apache-2.0
FlowBench Benchmark

A benchmark for evaluating embodied cognition systems end to end — measuring how errors in one stage propagate through the full loop rather than scoring each stage in isolation.

Benchmark suitePerception / Planning / ControlLicense: MIT
ManipNet Dataset

A dataset of manipulation trajectories recorded on hardware across insertion, assembly, and reorientation tasks, with per-frame contact and proprioceptive annotations for training world models.

Trajectory datasetContact annotationsLicense: CC-BY-4.0
sim2real-kit Software

A domain-randomisation and adversarial-distillation toolkit for training policies in simulation and transferring them to hardware with minimal task-specific tuning.

PythonIsaac Gym / MuJoCoLicense: MIT
All releases are available through our public repositories and are maintained as best-effort research code — they carry no production support guarantee. We welcome issues and reproductions, and we log any failure we can reproduce alongside the results it affects.