Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: mit
|
| 3 |
+
tags:
|
| 4 |
+
- reinforcement-learning
|
| 5 |
+
- panda-gym
|
| 6 |
+
- stable-baselines3
|
| 7 |
+
- sb3-contrib
|
| 8 |
+
- hw3
|
| 9 |
+
---
|
| 10 |
+
|
| 11 |
+
# HW3 — model checkpoints
|
| 12 |
+
|
| 13 |
+
Trained checkpoints for *EN.601.495/695 Introduction to Robot Learning, Spring 2026, HW3*.
|
| 14 |
+
|
| 15 |
+
Each subdirectory mirrors `starter-code/logs/<algo>/<env>_N/` from
|
| 16 |
+
[github.com/tarcode2004/hw3-rl](https://github.com/tarcode2004/hw3-rl)
|
| 17 |
+
and contains the `best_model.zip` (saved by SB3's `EvalCallback`),
|
| 18 |
+
the `evaluations.npz` curves, and the run's monitor csv.
|
| 19 |
+
|
| 20 |
+
The headline result is **TQC + HER + TimeFeatureWrapper on
|
| 21 |
+
PandaPickAndPlace-v3**: 92–98 % success on 50 deterministic eval
|
| 22 |
+
episodes. See the [standalone repo](https://huggingface.co/tarmus/tqc-PandaPickAndPlace-v3)
|
| 23 |
+
for that model.
|
| 24 |
+
|
| 25 |
+
## Layout
|
| 26 |
+
|
| 27 |
+
| Path | Algo / wrapper | Env | Notes |
|
| 28 |
+
|------|----------------|-----|-------|
|
| 29 |
+
| `zoo3/sac-PandaReach-v3` | zoo3 | sac-PandaReach-v3 | SAC + HER, sparse, converged 100% by 5k steps |
|
| 30 |
+
| `zoo3/sac-PandaPush-v3` | zoo3 | sac-PandaPush-v3 | SAC + HER, sparse, killed at 639k by reboot, best_model from peak |
|
| 31 |
+
| `zoo3/tqc-PandaPickAndPlace-v3` | zoo3 | tqc-PandaPickAndPlace-v3 | TQC + HER + TimeFeatureWrapper, sparse, 92-98% deterministic eval |
|
| 32 |
+
| `zoo3/ppo-PandaReach-v3` | zoo3 | ppo-PandaReach-v3 | PPO sparse, converged ~100K |
|
| 33 |
+
| `minimal_sac/PandaReachDense-v3` | minimal_sac | PandaReachDense-v3 | vanilla SAC, dense reward, 20K steps |
|
| 34 |
+
| `minimal_sac/PandaPickAndPlaceDense-v3` | minimal_sac | PandaPickAndPlaceDense-v3 | vanilla SAC, dense reward, 300K steps |
|
| 35 |
+
| `minimal_sac/PandaPushDense-v3` | minimal_sac | PandaPushDense-v3 | vanilla SAC, dense reward, 300K steps |
|
| 36 |
+
| `mbrl/PandaReachDense-v3` | mbrl | PandaReachDense-v3 | Basic MBRL (random data + dynamics + SAC on surrogate) |
|
| 37 |
+
| `mbrl/PandaPickAndPlaceDense-v3` | mbrl | PandaPickAndPlaceDense-v3 | Basic MBRL, partial run (stopped ~220K) |
|