---
pretty_name: ForceBenchmark WeightSorting V6
tags:
  - robotics
  - imitation-learning
---

# WeightSorting V6 / `ForcebenchmarkPandaWeightSorting-v2`

This release contains the 1,000 successful, physically paired demonstrations
used for the 100k-step diffusion-policy experiments. There are 500 scene pairs:
one light and one heavy rollout per scene. The task uses four masses (0.2, 0.4,
0.8, and 1.0 kg), balanced at 250 episodes each. The pair-disjoint split has
450 training pairs and 50 validation pairs. The reported 100k training jobs
used all 500 pairs; the split is supplied for future analysis.

Dataset: [HanLinqi/ForceBenchmark-WeightSorting-v6](https://huggingface.co/datasets/HanLinqi/ForceBenchmark-WeightSorting-v6).

## Files

| HF path | Contents | SHA-256 of H5 |
| --- | --- | --- |
| `planning_data/forcebenchmarkpandaweightsorting_v2/trajectory_rgb.h5` | 18D state (`qpos+qvel`), RGB, action, causal wrist wrench, auxiliary labels | `e17010001cb5e47b64e84b93b9b371f45c96015bf74a2c73265459b68920227b` |
| `planning_data_26d/forcebenchmarkpandaweightsorting_v2/trajectory_rgb.h5` | Same rollouts with 26D state (`qpos+qvel+TCP+gripper`) | `1c10151e66cf3f8aadd9ddc4b8d5e0c6637547f36b15564a20f842ff8fe1f5d4` |

Each H5 has a matching `trajectory_rgb.json`. The state18 directory also has
`pair_splits.json`, `decision_windows.json`, and `provenance/` with the raw
collection summary, audit and gate. `phase_id`, `stage_id`, scene
pair IDs, mass and route labels are auxiliary targets or audit metadata; they
are not policy observations. The raw audit state is not in either H5.

## Task and data construction

Import `task_force.forcebenchmark_panda_weight_sorting_v2` to register the
environment. The collector is
`task_force/scripts/generate_weight_sorting_v2_physical_pairs_v6.py`; its
campaign driver, audit, merge and conversion scripts are in `task_force/scripts`.
See `task_force/docs/weightsorting_v6_physical_pairs.md` for the full task
definition, collection protocol and physical assumptions.

The V2 action is a 7D increment to a persistent Cartesian reference. It is
incompatible with V1's instantaneous-TCP action semantics. The task uses an
ideal rigid grasp latch after physical finger contact and finite-precision
synthetic policy sensors; it does not represent ordinary friction-only
pinching or raw simulator proprioception. The same online sensors and
controller must be used for force and no-force policies. Online wrist wrench
settings are `force_torque_source="incoming_joint"`,
`force_substep_mean=True`, and `zero_force_on_reset=True`.

Collection audited exact shared pre-decision actions and reference positions,
causal wrench, sensor replay, physical pair separation and five-fold
scene-grouped leakage probes. Collection accepted 1,000 episodes/500 pairs.
The separate pilot used evaluation seeds 0–4, 20 episodes per seed. For each
trained experiment, use five evaluation seeds ×20 episodes per seed, with
evaluation scenes separate from collection scenes.