DP3 (Diffusion Policy 3D)
DP3 is a 3D-aware diffusion policy: it consumes a sparse point cloud of the workspace instead of a 2D image, and samples actions via iterative denoising. Strong on the fine-grained / contact-rich tasks that motivate MetaFine — peg-in-hole, plug-in-socket, lid-opening.
What it is
DP3 (3D Diffusion Policy) — a diffusion policy that conditions on point-cloud observations. Vendored under core/policies/diffusion_policy_3d.
Data
DP3 needs point-cloud observations: replay with a depth/pointcloud obs mode, then convert. See Quickstart for the record → merge → replay → convert flow.
Install & run
$ pip install -e core/policies/diffusion_policy_3d
DP3's training / eval entrypoints and flags are documented in its own in-repo guide:
$ cat core/policies/diffusion_policy_3d/README.md
Status: vendored. The package path is core/policies/diffusion_policy_3d (not dp3); treat its README as the source of truth.