File size: 1,601 Bytes
1754c73
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
---
license: cc-by-4.0
tags: [robotics, grasping, rag, force-semantic-map, heft]
pretty_name: HEFT force-semantic map (5,278 real products)
---
# HEFT force-semantic map

The retrieval memory ("RAG map") of **HEFT** (Hierarchical Estimation + Feedback for Tactile-free grasping, Correll Lab, MAGPIE gripper). Code: [github.com/ACWArob/heft](https://github.com/ACWArob/heft) (private).

Each entry is one real product with:
- DINOv2-small photo embedding + log dimensions (`heft.perception.embedding_dims`)
- physical properties: mass (listed weight), friction, stiffness, each with a log-space sd and a source (`listed`, `gemini`, later `measured`)
- a reasoning description (product type + material, e.g. "drinking cup, plastic") and its text embedding (`paraphrase-multilingual-MiniLM-L12-v2`), used by `ForceSemanticMemory.lookup_text`
- volume, so mass transfers between objects as density x volume

5,278 products from Amazon Berkeley Objects (5–800 g, smallest side ≤ 10 cm) + 37 entries with Gemini friction/stiffness. On held-out products, the image lookup is within 1.26× of the listed weight with 4,000 entries in the map (1.08× for its most confident quarter).

The map is meant to grow on the robot: every grasp writes a measured mass for that object, with its description, and measured entries outweigh listed or generated ones. No retraining.

```python
from heft import ForceSemanticMemory
mem = ForceSemanticMemory("abo_force_map.json.gz", k_nn=5, temp=0.15)
```

Source data: Amazon Berkeley Objects (Collins et al., CVPR 2022), CC BY 4.0. Listed weights can include packaging.