tOMRmA commited on
Commit
b9dce0a
·
verified ·
1 Parent(s): f836c19

Add checkpoint documentation

Browse files
Files changed (1) hide show
  1. README.md +62 -3
README.md CHANGED
@@ -1,3 +1,62 @@
1
- ---
2
- license: bsd-3-clause
3
- ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # TacInsert Checkpoints
2
+
3
+ This directory is reserved for optional pretrained checkpoints used for quick simulation smoke tests. Checkpoint files are large binary artifacts and are intentionally ignored by git. For public releases, upload the `.pth` files to GitHub Releases, Hugging Face, or Zenodo, and keep this README plus checksums in the source repository.
4
+
5
+ ## Available Local Checkpoints
6
+
7
+ The following checkpoints were used to validate the current TacInsert simulation code:
8
+
9
+ | File | Task | Observation dim | Notes | SHA256 |
10
+ | --- | --- | ---: | --- | --- |
11
+ | `TacInsert-LHole-III-Direct-v0.pth` | `TacInsert-LHole-III-Direct-v0` | 23 | L-hole Tol. III with contact-force observation and `fixed_sigma: True` | `FC158B0D72AFDAB0D860CEB379F8E3EC2B0B693D49C9B24629608A5BEA4B58AC` |
12
+ | `TacInsert-Manipulation-Square-SingleHole-Direct-v0.pth` | `TacInsert-Manipulation-Square-SingleHole-Direct-v0` | 26 | ManipulationNet-style square/rectangle single-hole sampler policy with contact-force observation, tolerance one-hot, and `fixed_sigma: False` | `0301CD4C57848EDA73FAE2D9F019D879C6D8AA1F29315CCA8A12BDC5B36CA10E` |
13
+
14
+ ## Smoke-Test Commands
15
+
16
+ Run from the repository root after installing the TacInsert extension.
17
+
18
+ ### L-Hole Tol. III
19
+
20
+ ```powershell
21
+ python scripts/rl_games/play.py `
22
+ --task TacInsert-LHole-III-Direct-v0 `
23
+ --num_envs 128 `
24
+ --headless `
25
+ --checkpoint source/TacInsert/TacInsert/tasks/direct/tacinsert/checkpoints/TacInsert-LHole-III-Direct-v0.pth
26
+ ```
27
+
28
+ Expected behavior: the terminal prints one summary per completed episode. In local validation with 128 environments, three episodes reached approximately 94-98 percent success.
29
+
30
+ ### ManipulationNet-Style Square Single-Hole
31
+
32
+ ```powershell
33
+ python scripts/rl_games/play.py `
34
+ --task TacInsert-Manipulation-Square-SingleHole-Direct-v0 `
35
+ --num_envs 128 `
36
+ --headless `
37
+ --checkpoint source/TacInsert/TacInsert/tasks/direct/tacinsert/checkpoints/TacInsert-Manipulation-Square-SingleHole-Direct-v0.pth
38
+ ```
39
+
40
+ Expected behavior with the default fixed Tol. I sampler: the terminal prints one summary per completed episode. In local validation with 128 environments, three episodes reached approximately 98-99 percent success.
41
+
42
+ To evaluate another fixed tolerance, edit `multi_hole_sample_weights` in `TacInsertManipulationSquareSingleHole` before running playback. For example, Tol. IV evaluation uses:
43
+
44
+ ```python
45
+ multi_hole_sample_weights = [0.0, 0.0, 0.0, 1.0]
46
+ ```
47
+
48
+ In local Tol. IV validation with 128 environments, three episodes reached approximately 61-63 percent success.
49
+
50
+ ## Contact-Force Logging Check
51
+
52
+ For tasks with `contact_force["log_contact_force"] = True`, CSV logging is enabled only in evaluation mode with `--num_envs 1`:
53
+
54
+ ```powershell
55
+ python scripts/rl_games/play.py `
56
+ --task TacInsert-Manipulation-Square-SingleHole-Direct-v0 `
57
+ --num_envs 1 `
58
+ --headless `
59
+ --checkpoint source/TacInsert/TacInsert/tasks/direct/tacinsert/checkpoints/TacInsert-Manipulation-Square-SingleHole-Direct-v0.pth
60
+ ```
61
+
62
+ CSV files are written to `contact_force_logs/`, which is ignored by git.