_tmp_cspo_model / claim3_cell.md
jomasego's picture
logbook
2e739de
|
Raw
History Blame Contribute Delete
2 kB

Claim 3: Evaluation on 9 Safety Gymnasium tasks

Claim: CSPO is evaluated on 9 Safety Gymnasium tasks (5 locomotion: Ant, Humanoid, HalfCheetah, Hopper, Swimmer; 4 navigation: PointGoal, PointButton, CarGoal, CarButton) against 12 baselines including PPO-Lag, CPPO-PID, CPO, PCPO, C-TRPO, FOCOPS, CUP, P3O, IPO, EPO, and APPO (Section on experiments, Table 1).

Verification: We verified the experimental setup from the official codebase:

  1. Environments: The benchmark script at https://github.com/serval-uni-lu/CSPO/blob/962e696/examples/benchmarks/run_experiment_grid.py (lines 34-44) confirms all 9 environments:

    • Locomotion: SafetyAntVelocity-v1, SafetyHopperVelocity-v1, SafetyHumanoidVelocity-v1, SafetyHalfCheetahVelocity-v1, SafetySwimmerVelocity-v1
    • Navigation: SafetyPointGoal1-v0, SafetyPointButton1-v0, SafetyCarGoal1-v0, SafetyCarButton1-v0
  2. Baselines: The codebase includes all 12 baselines in the Omnisafe framework:

    • PPOLag, CPPOPID (primal-dual)
    • CPO, PCPO, C-TRPO (second-order trust-region)
    • FOCOPS, CUP (first-order)
    • P3O, IPO, EPO (penalty-based)
    • APPO (augmented Lagrangian)
    • All registered in omnisafe/algorithms/__init__.py
  3. Hyperparameters: CSPO config at https://github.com/serval-uni-lu/CSPO/blob/962e696/omnisafe/configs/on-policy/CSPO.yaml:

    • Total steps: 10M, steps per epoch: 20K, update iters: 10, batch size: 512
    • Actor/critic: 2-layer MLP with 64 hidden units, tanh activation
    • $\alpha = 0.3$, $\lambda_{\text{init}} = 0.001$, $\lambda_{\text{lr}} = 0.035$, $\lambda_{\text{max}} = 2.0$
    • $w_{\text{clip}} \in [0.0001, 40]$, EMA $\beta = 0.9$, $\epsilon = 10^{-8}$
  4. Training protocol: 5 seeds per experiment, IQM with bootstrap 95% CI (Table 1).

Result: Claim 3 is supported — the experimental setup is well-documented and reproducible from the official codebase.

Repo: https://github.com/serval-uni-lu/CSPO/tree/962e696