CrocoWilly/geometry-conditioned-insertion

Geometry-Conditioned Insertion

Human teleoperation demonstrations for a peg-in-hole insertion benchmark that scores a policy on its behavior during the rollout, not only on whether it finished the task.

Collected in NVIDIA Isaac Lab with a Franka Panda. 420 episodes across six datasets, at least 50 per experiment.

Code, benchmark and results: https://github.com/CrocoWilly/do-policies-adapt-to-geometry

Accepted at IEEE/SICE SII 2027, Kobe, Japan.

---

What the experiments measure

Five independent experiments, each isolating a different aspect of geometric capability. Every dataset below is the training data for one of them.

File	Episodes	Size	Experiment	Capability under test
Diagnosis0_dataset_forTraining.hdf5	50	1.1 GB	Skill Check	Can the policy insert at all?
Diagnosis1_dataset_forTraining.hdf5	50	1.1 GB	Alignment Adaptation	Does it align more for a tight hole than a loose one?
Diagnosis2_dataset_forTraining.hdf5	50	1.3 GB	Strategy Diversity	When two orientations both fit, does it use both?
Diagnosis3_dataset_forTraining.hdf5	60	1.6 GB	Strategy Disambiguation	When only one fits, does it pick that one?
Diagnosis4_dataset_forTraining.hdf5	90	1.9 GB	Hole Identification	Does it match peg shape to hole shape?
MainExp_dataset_forTraining.hdf5	120	2.6 GB	Appearance × geometry sweep	Used for the auxiliary-loss experiments
MainExp_dataset_anchorPool_forTraining.hdf5	2,578 anchors	304 MB	—	Anchor pool for the Relative Geometry-Appearance Loss

The Diagnosis sets are independent, so there is no need to download all 10.1 GB to run one experiment.

Download

pip install huggingface_hub

# everything
hf download CrocoWilly/geometry-conditioned-insertion \
    --repo-type dataset --local-dir datasets/

# or a single experiment
hf download CrocoWilly/geometry-conditioned-insertion \
    Diagnosis1_dataset_forTraining.hdf5 \
    --repo-type dataset --local-dir datasets/

Format

Standard robomimic HDF5 layout — a data group with one subgroup per episode.

data/                     attrs: env_args
  demo_0/                 attrs: num_samples, success
    obs/                  19 keys, 22 in three of the datasets — see below
    actions           (T, 4)  float32  the key robomimic trains on
    action_world_4d   (T, 4)  float32  [dx, dy, dz, dyaw], world frame
    action_hole_4d    (T, 4)  float32  the same action in the hole frame
    processed_actions (T, 8)  float32  the raw recorded action, before conversion
    states/                            full simulator state, for replay
    initial_state/                     episode reset state
  demo_1/
  ...

The action space is 4D. Roll and pitch are held fixed and the gripper stays closed on the peg, so the policy controls exactly the degrees of freedom the insertion needs and nothing else. actions is a byte-identical copy of action_world_4d, kept under the name robomimic expects.

states and initial_state are retained, so an episode can be replayed in Isaac Lab rather than only used as supervision.

Observations

Every episode records far more than the baselines consume. The published BC-LSTM and Diffusion Policy baselines train on three keys only — marked [train] below. The rest are recorded so the data supports experiments beyond the ones in the thesis: depth-based policies, wrist-camera policies, joint-space control, or privileged-information baselines.

Key	Shape (per step)
table_side_cam	(200, 200, 3) uint8	[train] side-view RGB
ee_pos_world	(3,) float32	[train] end-effector position, world frame
ee_yaw_world	(1,) float32	[train] end-effector yaw, world frame
table_front_cam	(200, 200, 3) uint8	front-view RGB
wrist_cam	(200, 200, 3) uint8	wrist-mounted RGB
table_side_cam_depth	(200, 200, 1) float32	side-view depth
table_front_cam_depth	(200, 200, 1) float32	front-view depth
wrist_cam_depth	(200, 200, 1) float32	wrist-mounted depth
joint_pos, joint_vel	(9,) float32	arm + gripper joint state
ee_quat_world	(4,) float32	end-effector orientation
peg_pos_world, peg_quat_world	(3,), (4,) float32	peg pose
hole_pos_world, hole_quat_world	(3,), (4,) float32	hole pose
hole_yaw_world	(1,) float32	hole yaw
ee_pos_rel, ee_yaw_rel	(3,), (1,) float32	end-effector pose in the hole frame
clearance	(1,) float32	peg-hole clearance for this episode

Those 19 keys are present in every dataset. Three of them carry extra keys for what their experiment varies:

Dataset	Extra keys
Diagnosis3	ee_yaw_rel_slot2, peg_pos_rel, peg_yaw_rel — the second candidate slot
Diagnosis4	peg_shape_no, hole_shape_codes (3,), hole_slot_positions_world (9,) — which of three holes matches the peg
MainExp	box_color_no, peg_texture_no, table_texture_no — the appearance factors

Because geometry and appearance are recorded separately per episode, pairs of observations differing in exactly one of the two can be sampled — which is what the Relative Geometry-Appearance Loss needs.

Anchor pool

A separate file holding 2,578 decision-critical time-steps — the peg close to the hole surface and still misaligned — sampled by the auxiliary loss. Built from MainExp_dataset_forTraining.hdf5.

anchors/
  anchor_0/
    obs/
      table_side_cam   (200, 200, 3)
      ee_pos_world     (3,)
      ee_yaw_world     (1,)
    attrs: color, peg_texture, table_texture, size, delta_z, yaw_error_deg
  ...
attrs: source_file, total_anchors, z_diff_min, z_diff_max,
       yaw_error_threshold_deg, yaw_rotation_diff_threshold,
       peg_length, hole_height

Collection

Keyboard teleoperation in Isaac Lab, recorded at 25 Hz. Teleoperation was retuned for insertion rather than pick-and-place: position sensitivity lowered to 0.0125 for the fine alignment the task demands, rotation sensitivity raised so yaw correction stayed practical.

The recorder applies per-episode randomization of geometry and appearance, so each episode is a different instance of its experiment rather than a repeat.

Raw recordings use a 7D action; the published files have been converted to the 4D form and had per-episode clearance and depth scaling attached. The conversion scripts are in the GitHub repository.

Inspecting a file

import h5py

with h5py.File("Diagnosis1_dataset_forTraining.hdf5", "r") as f:
    demos = list(f["data"].keys())
    print(len(demos), "episodes")
    d = f["data"][demos[0]]
    print({k: d["obs"][k].shape for k in d["obs"]})
    print("actions", d["actions"].shape)

The repository also ships Check_hdf5_structure.py, List_episodes.py and View_hdf5_images.py for structure dumps and frame-by-frame playback.

Citation

The benchmark and the relative loss are published at IEEE/SICE SII 2027:

@inproceedings{chen2027geometry,
  title     = {Do Robot Policies Adapt to Geometry? Diagnosing and Learning
               Geometry-Conditioned Behavior in Insertion Tasks},
  author    = {Chen, Yen-Chang and Chu, Wei-Ta},
  booktitle = {Proceedings of the IEEE/SICE International Symposium on
               System Integration (SII)},
  address   = {Kobe, Japan},
  year      = {2027},
  note      = {To appear}
}

The full study, including the experiments not covered in the paper:

@mastersthesis{chen2026geometry,
  title  = {Do Robot Policies Adapt to Geometry? Diagnosing and Learning
            Geometry-Conditioned Behavior in Insertion Tasks},
  author = {Chen, Yen-Chang},
  school = {National Cheng Kung University},
  year   = {2026}
}

Acknowledgment

Collected at the Multimedia and Computer Vision Laboratory, National Cheng Kung University, under the supervision of Prof. Wei-Ta Chu.

Funded in part by the National Science and Technology Council, Taiwan, under grants 115-2622-8-006-015, 114-2622-E-006-028, 114-2221-E-006-047-MY3, 112-2221-E-006-136-MY3, 115-2425-H-006-005, and 114-2634-F-006-002.

License

BSD-3-Clause, matching Isaac Lab. The demonstrations were collected by the author in simulation; no third-party or human-subject data is involved.