Safety-RL

Benchmarks

Env	PPO		PPO_Lagrangian		PPO_CRPO		PPO_CRPO_SR
	Return	Cost	Return	Cost	Return	Cost	Return	Cost
SafetyHalfCheetahVelocity-v1	4927.870±1265.715	823.125±219.088	2074.117±748.714	123.192±69.233	608.081±425.514	14.582±17.990	611.925±259.232	0.032±0.203
SafetySwimmerVelocity-v1	43.901±18.263	36.776±29.344	-4.965±5.780	0.321±2.787	15.633±3.435	34.992±14.742	17.970±6.626	2.469±5.474
SafetyPointGoal1-v0	11.829±6.338	95.632±112.685	-9.743±8.966	10.093±34.616	23.726±3.107	64.762±43.774	7.047±4.763	32.575±45.225

Supported algorithms

1. PPO : Proximal Policy Optimization Algorithms
2. PPO_Largrangian : Benchmarking Safe Exploration in Deep Reinforcement Learning
3. PPO_CRPO : CRPO: A New Approach for Safe Reinforcement Learning with Convergence Guarantee
4. PPO_CRPO_SR : SAFETY REPRESENTATIONS FOR SAFER POLICY LEARNING