Blame for README.md

---
license: other
task_categories:
- robotics
tags:
- simulation
- isaac-sim
- lab-automation
- microplate
- imitation-learning
---

# MagicBioLab — microplate cell (UR5e + plate gripper), simulated demonstrations

Scripted demonstrations of a full lab-automation cycle in a simulated bio-lab enclosure:
**conveyor carrier → plate hotel → a4S plate sealer → back onto the carrier**, with a UR5e arm and a
parallel plate gripper. Every episode is one complete, successful cycle on one microplate.
Data are added continuously while collection runs; `index/episodes.csv` lists every uploaded episode.

## Layout

```
episodes/<episode>/
  data.npz               per-frame arrays (25 Hz)
  wrist.mp4  env.mp4     640x480, 25 fps, frame-aligned with data.npz
  meta.json              frame count, joint order, frames used, dr_version
  scene_snapshot.json    full scene state at episode start (all DR values: lights, poses, belt speed, cameras)
  device_events.jsonl    conveyor / a4S device events (station arrival, drawer, seal, release)
index/episodes.csv       one row per episode: DR parameters, plate type, length, checks, dr_version
```

## data.npz (T frames, 25 Hz; robot base frame)

| key | shape | meaning |
|---|---|---|
| joint_pos / joint_vel | (T, 8) | 6 arm joints + 2 finger joints (m) |
| eef_pos / eef_quat | (T, 3) / (T, 4) | gripper pose, quat wxyz |
| gripper_pos | (T, 2) | finger positions |
| camera_pose | (T, 2, 7) | local camera pose (wrist: in the gripper base link; env: in the env frame), pos + quat wxyz |
| action_joint_pos | (T, 8) | **joint-position action**: commanded joint targets for the next step |
| action_joint_delta | (T, 8) | action_joint_pos − joint_pos |
| action_eef_pose | (T, 7) | **EEF-pose action**: forward kinematics of action_joint_pos (pos + quat wxyz) |
| action_eef_delta | (T, 6) | EEF delta (translation + rotation vector) |
| action_gripper | (T, 1) | finger target (m), 0.018 = open |
| movel_goal | (T, 8) | the scripted planner's current Cartesian goal (NaN when none) |
| tick / phase / caption | (T,) | 50 Hz control tick, task phase (12 phases), language caption |

Open gripper = 0.018 m per finger. The simulation runs a 50 Hz control loop; every other tick is kept.

## Cameras

* **wrist** — on the plate gripper, looking past the fingers. Per-episode mount DR (±3 mm, ±1.5°),
  per-frame shake (OU jitter ~1.2 mm / 0.35°) and sway from the arm's acceleration.
* **env** — top-front of the enclosure, looking down at the cell. Per-episode pose DR (±2.5 cm, ±2°); static within an episode.

## Domain randomization (per episode)

Lighting (ceiling rect light + a free sphere/disk key light: pose, intensity, colour temperature, colour;
dome intensity) · 16 microplate types (SBS 6–1536 well, PCR, deep-well, reservoir; clear/black/white/coloured) ·
plate offset (±15 / ±12 mm) and yaw (±12°) on the carrier · carrier stop position at the pick station ·
belt speed (0.15–0.30 m/s) · arrival time of the carrier · a4S sealer pose (7 × 8 cm, ±5°) ·
seal dwell and hand-off timing · force/torque pulses (5–20 N) on the held plate during every transfer
(grip-stability test; the episode is discarded if the plate is lost) · camera poses (above).

`dr_version`: **v2** episodes re-draw the wrist-camera mount and the env-camera pose every episode;
**v1** episodes (the first ~1050) drew both camera poses once per simulator process and environment
(116 distinct values among the first 951). In all versions the a4S sealer pose is drawn once per
simulator process and environment (4 values per process). Columns `cam_dr_per_episode` /
`a4s_dr_per_episode` in `index/episodes.csv` give the flags per episode. All other DR is per episode.

## Quality gates (every uploaded episode)

Task succeeded; video frame count = data frame count for both cameras; no NaN in state/action arrays;
all 12 phases present and ending in `release`; env camera static within the episode.
Replay of recorded actions reproduces the joint trajectories (mean 0.0003 rad) and final plate pose (2.4 mm).