ParamTatva/mujoco-sota-benchmark

MuJoCo SOTA Benchmark

Standard MuJoCo continuous control benchmarks from Gymnasium used to evaluate reinforcement learning algorithms.

Benchmark Environments

Environment	Obs Dim	Act Dim	CleanRL SOTA	ParamTatva Best
Hopper-v5	11	3	2,382 +/- 271	3,183.2 (134%)
Walker2d-v5	17	6	~4,000	4,918.5 (123%)
HalfCheetah-v5	17	6	~6,000	5,803.9 (97%)
Reacher-v5	8	2	~-4	-4.2 (~100%)
Ant-v5	27	8	~5,000	886.6 (training)
Humanoid-v5	348	17	~5,000	573.8 (training)

How to Evaluate

pip install torch gymnasium[mujoco]

import torch
import gymnasium as gym

# Load checkpoint
checkpoint = torch.load("hopper_v5_sota.pt")
agent = Agent(obs_dim=11, act_dim=3)
agent.load_state_dict(checkpoint["model_state_dict"])

# Evaluate
env = gym.make("Hopper-v5")
returns = []
for ep in range(100):
    obs, _ = env.reset()
    total = 0
    done = False
    while not done:
        action = agent.get_action(torch.FloatTensor(obs))
        obs, reward, term, trunc, _ = env.step(action.detach().numpy())
        total += reward
        done = term or trunc
    returns.append(total)

print(f"Mean: {sum(returns)/len(returns):.1f}")

Reference

• Model: ParamTatva/sanskrit-ppo-hopper-v5
• Blog: The 371 Wall
• CleanRL Baselines: vwxyzjn/cleanrl

License

Apache 2.0

ParamTatva.org 2026