Vladprime/td3-PandaReachDense-v3

TD3 + HER — PandaReachDense-v3

Trained from random initialization for Vlad Kolomiiets (Vladprime) for HF Deep RL Unit 6. TD3 uses hindsight experience replay with future-goal relabeling, the original dense environment reward and original observation/action spaces. Training seed 20261101; 10000 environment steps. Separate validation seeds selected the checkpoint.

Independent evaluation: 200 fixed seeds 1000000–1000199 after reloading the checkpoint. Mean reward -0.2152, standard deviation 0.1161, mean minus standard deviation -0.3313; final-step success rate 100.0%.

Reproduce

Use Python 3.10 on Linux (the pinned NumPy version predates Python 3.12).

pip install -r requirements.txt
python train.py

from stable_baselines3 import TD3
import gymnasium as gym
import panda_gym
model=TD3.load("td3-PandaReachDense-v3.zip",env=gym.make("PandaReachDense-v3"))