angledusgar commited on
Commit
447a610
·
verified ·
1 Parent(s): 6b6b2c3

Forge peginsert PPO expert (300 epochs)

Browse files
Files changed (2) hide show
  1. README.md +50 -0
  2. forge_peginsert_2026-06-30.pth +3 -0
README.md ADDED
@@ -0,0 +1,50 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ tags:
4
+ - robotics
5
+ - reinforcement-learning
6
+ - isaac-lab
7
+ - forge
8
+ - rl-games
9
+ - ppo
10
+ ---
11
+
12
+ # Forge PegInsert expert (PPO)
13
+
14
+ PPO policy for `Isaac-Forge-PegInsert-Direct-v0` ([Isaac Lab](https://github.com/isaac-sim/IsaacLab) FORGE suite),
15
+ trained with `rl_games`. Final return **345.5**; measured success **~96% over ~1,010 collection episodes**.
16
+
17
+ ## ⚠️ This policy is for **stock** PegInsert only
18
+
19
+ It was trained on 2026-06-30, before the task config in our Isaac Lab checkout was modified for a
20
+ separate study (spawn raised to 8 cm with ±5 cm lateral noise, and a ±10° tilt of the peg *inside
21
+ the gripper*). On stock dynamics it succeeds ~96% of the time; on that modified variant it scores
22
+ **0/12**. A policy trained from scratch on the modified variant plateaus at a return of 72.7 against
23
+ this one's 345.5 — the geometry explains why: with a 7.986 mm peg in an 8.100 mm hole, a peg cocked
24
+ beyond ~9.6° cannot enter at all, and the tilt is not observable in the policy's inputs.
25
+
26
+ On an unmodified Isaac Lab, run it as-is.
27
+
28
+ ## What is in the file
29
+
30
+ `rl_games` checkpoints hold more than the policy: the actor-critic weights, the optimizer state,
31
+ and — importantly for inference — the running observation normalisation statistics. That is why a
32
+ small MLP policy takes 204 MB. Load it through the Isaac Lab player rather than `torch.load` alone,
33
+ so the normaliser is restored with the weights:
34
+
35
+ ```bash
36
+ ./isaaclab.sh -p scripts/reinforcement_learning/rl_games/play.py \
37
+ --task Isaac-Forge-PegInsert-Direct-v0 --checkpoint /path/to/forge_peginsert_2026-06-30.pth --num_envs 1 --headless
38
+ ```
39
+
40
+ ## Training
41
+
42
+ Stock Isaac Lab `rl_games` PPO, unchanged hyper-parameters except the epoch count:
43
+ 128 environments, seed 0, **300 epochs** (the shipped config stops at 200, which cuts the run while
44
+ the return is still climbing). Snapshots exist at epochs 100 and 200; this repo ships epoch 300.
45
+
46
+ ## Related
47
+
48
+ - Demonstrations produced with this policy: [`angledusgar/forge-v1`](https://huggingface.co/datasets/angledusgar/forge-v1)
49
+ - Its failures: [`angledusgar/forge-failure-v1`](https://huggingface.co/datasets/angledusgar/forge-failure-v1)
50
+
forge_peginsert_2026-06-30.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:10efdf802e8abfccaa13f6e7a18cc2887ca2f3f96b35526bc9984fde419e59f8
3
+ size 213431615