tags:
- robotics
- dexterous-manipulation
- bimanual
- pose-estimation
- diffusion-policy
pretty_name: DP-T perceived-pose datasets (bimanual_assembly)
DP-T perceived-pose datasets (bimanual_assembly)
These are training data for perceived-pose policies: the 100 DexJoCo bimanual_assembly demos, with obj_pose replaced by the poses our perception estimated at every step of every demo.
The checkpoints trained on them are in b12902136/DP-T-perceived-pose. The code is the DP-T repository, branch cadfree-pointflow; how these were made is in docs/PERFECT_DEPTH_PIPELINE.md.
| Folder | Contents | Size |
|---|---|---|
rand_obj_objfut_pointflow_perceived_dpe/bimanual_assembly |
dpe: depth tracker on perfect (simulator) depth, every step, new canonical rotations, mesh frame | 31 MB |
rand_obj_objfut_pointflow_perceived_rpe/bimanual_assembly |
rpe: the same run's RGB tracker poses | 32 MB |
every_step_perfect/tracks |
the per-step depth-tracker poses (one json per demo) that dpe was built from | 40 MB |
every_step_perfect/tracks_rgb |
the per-step RGB-tracker poses that rpe was built from | 75 MB |
every_step_perfect/lite |
per-frame camera poses, robot state and true object poses (no images); the dataset builder's --frames |
16 MB |
Dataset contents (replay_buffer.zarr, 54,170 steps, 100 episodes)
| Array | What it is |
|---|---|
data/obj_pose |
The policy input: the perceived pose [tray 7, peg 7], position + quaternion (wxyz), in the frame of the shared SAM 3D meshes, at every step (no hold, no interpolation) |
data/obj_pose_fut |
The pointflow target: the TRUE motion applied to the perceived pose. Ground truth, used as supervision only. |
data/obj_pose_gt |
The true poses, for measurement only. No training config reads it. |
data/state, data/action, data/timestamp, meta/episode_ends |
Unchanged from the source dataset rand_obj_objfut_pointflow |
pointflow_stats.npz holds the pointflow statistics on the perceived cloud. perceived_pose.json holds the builder's settings: --canon --peg-spin hand, frame mesh, every step.
Rotations. Rotations are canonicalised:
- the tray's yaw is folded onto one symmetry branch;
- the peg axis points toward the right hand;
- the peg's spin is taken from the right TCP.
The eval perception must use the same rule (--perceived-canon --perceived-peg-spin hand).
Perception error against the true pose, as median / p90:
| Tray | Peg | |
|---|---|---|
| dpe | 1.62 / 2.2 mm | 1.10 / 2.3 mm |
| rpe | 1.86 / 6.2 mm | 1.47 / 3.8 mm |
Use
# the datasets, into dexjoco's converted_datasets/zarr (where slurm/main reads PV=dpe / rpe)
hf download b12902136/DP-T-perceived-pose-data --repo-type dataset --include 'rand_obj_objfut_pointflow_perceived_*' \
--local-dir $DEXJOCO_ROOT/converted_datasets/zarr
# (optional) the per-step tracks, to rebuild other representations without re-running perception
hf download b12902136/DP-T-perceived-pose-data --repo-type dataset --include 'every_step_perfect/*' --local-dir $VLA_ROOT/DP-T/outputs/depth
# train (the code repo; also run assets/perceived_pose/install.sh once for the policy's object cloud)
PV=dpe VARIANT=N POSE=perceived SEED=42 sbatch slurm/main/train.sbatch
To rebuild a representation, run scripts/policy/add_perceived_pose_zarr.py with --tracks …/every_step_perfect/tracks (or tracks_rgb) and --frames …/every_step_perfect/lite, plus the source dataset rand_obj_objfut_pointflow. Give it a new tag, since the builder refuses an existing output.
Caveats
- Perfect depth. dpe's depth is the simulator's own, noise-free, so it is an upper bound for a depth sensor.
- Shared depth model. The tracker's shared object model was built from these same 100 training demos. It used only depth, with no CAD; the eval scenes were not involved.
- Assembly only. These datasets are specific to bimanual_assembly's objects and the DexJoCo demo set.