Datasets:
provenance: experiment line + what changed vs parent
Browse files
README.md
ADDED
|
@@ -0,0 +1,18 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
tags:
|
| 4 |
+
- LeRobot
|
| 5 |
+
- robotics
|
| 6 |
+
---
|
| 7 |
+
|
| 8 |
+
<!-- dreamer-provenance -->
|
| 9 |
+
## Experiment provenance
|
| 10 |
+
|
| 11 |
+
**Experiment line:** stack line (colour-cube world). Three cubes differing ONLY in hue; day cycles relabel grasp → pick up → stack. Run pages: https://storage.googleapis.com/rbhowmik-dreamer/site/stack/index.html
|
| 12 |
+
|
| 13 |
+
The full-scene, prompt-necessary corpus that solved hue binding (go over the <colour> cube 10/10): every scene holds all three cubes, so the colour word is the only thing that disambiguates.
|
| 14 |
+
|
| 15 |
+
**What changed:** vs `stack_v2`: episodes 2101 → 2100, frames 63143 → 62659. Same task vocabulary — the new episodes change coverage, not words.
|
| 16 |
+
|
| 17 |
+
Collected in simulation (robosuite/MuJoCo, Panda): scene + wrist video, proprio, OSC actions, and a language task string minted by the project's verbalizer (certificate → instruction; the relabeler may not invent words).
|
| 18 |
+
<!-- /dreamer-provenance -->
|