README.md

---
license: mit
task_categories:
- robotics
tags:
- robotics
- world-models
- isaac-sim
- sim2real
- manipulation
---

# UR10e Linear Gripper — Jig / Bottom Enclosure — SimDist Stage-2 Dataset

Action-conditioned simulation rollouts for pretraining a latent world model, generated with the
**Simulation Distillation (SimDist)** stage-2 procedure ([arXiv:2603.15759](https://arxiv.org/abs/2603.15759),
RSS 2026; code [CLeARoboticsLab/simdist](https://github.com/CLeARoboticsLab/simdist), MIT).

**Status: placeholder — generation not yet run.** This card describes the dataset that will land here.

## Task

Insertion of `JigV2` (insertive) into `BottomEnclosure` (receptive), on a UR10e with a custom linear
two-jaw gripper, simulated in IsaacLab. 10 Hz control, 16 s episodes (160 steps), 6-D relative
end-effector pose (Cartesian OSC) plus a binary gripper command.

## How it was generated

SimDist Algorithm 2. Per environment, a diagonal action-noise covariance is sampled; on each reset the
environment is reassigned to the expert policy with probability 0.5, else to a uniformly drawn
sub-optimal checkpoint. Gaussian action perturbations are injected in contiguous bursts of `U[1,5]`
steps interleaved with clean stretches of `U[5,10]` steps (the paper's *manipulation* intervals). The
binary gripper dimension is perturbed by flipping the commanded open/close rather than by additive
noise, since additive noise on a thresholded signal has no controllable effect.

The point of the perturbations is coverage: a planner searches far outside the expert distribution, so
the dataset must contain mistakes, recoveries and failures, not just clean expert trajectories. The
paper's own ablation shows expert-only data collapses success from 0.90 to 0.10.

Expert policy, checkpoint ladder and value function come from
[UR10e-LinearGripper-Jig-BottomEnclosure-Stage1](https://huggingface.co/buckets/RubetekRobotics/UR10e-LinearGripper-Jig-BottomEnclosure-Stage1);
reset states from
[UR10e-LinearGripper-Jig-BottomEnclosure-Resets](https://huggingface.co/datasets/RubetekRobotics/UR10e-LinearGripper-Jig-BottomEnclosure-Resets).

## Contents

One HDF5 file per shard. All arrays share a leading timestep axis.

| Field | Type | Notes |
|---|---|---|
| `proprio` | float32 `(N, 20)` | last gripper action (1), last arm action (6), arm joint positions (6), end-effector pose (6), binary contact (1) |
| `front_rgb`, `side_rgb`, `wrist_rgb` | vlen uint8 `(N,)` | JPEG-encoded, quality 90 |
| `actions` | float32 `(N, 7)` | raw policy output, **before** the environment's action scale |
| `rewards` | float32 `(N,)` | |
| `values` | float32 `(N,)` | V(s) from the expert critic, always the final checkpoint |
| `expert_flags` | bool `(N,)` | true only when on the expert **and** the action was uncorrupted |
| `env_ids`, `episode_ids` | int32 `(N,)` | episode boundary markers |

**Observations are deliberately non-privileged.** No ground-truth object pose appears anywhere in the
observation fields — a world-model encoder that could read object pose directly would defeat the entire
premise. Privileged simulator state is used only for the reward and the critic.

Images are stored JPEG-encoded rather than raw: raw uint8 is 451,584 B/step against ~30,714 B/step
encoded at 224 px, and a float32 representation would be 1.8 MB/step.

## Known limitations

- Dataset scale is the axis the method is most sensitive to (paper Table I: 0.90 at full scale, 0.72 at
  50 %, 0.06 at 10 %). The realised scale here is recorded below once generation completes.
- The expert was trained against a reset mixture in which `ObjectPartiallyAssembledEEGrasped` was only
  ~33 % genuinely holding the object, due to a recorder bug since fixed. That cannot be undone and is
  documented rather than hidden.

## Licence

MIT, following upstream SimDist.