HalfCheetah PPO — Standard Training (Before)

Deployment Robustness

FAIL

Return 2% (jitter / delay / spike)

Stress Robustness

FAIL

Return -12% (5x speed)

Deployment Fragile

Agent degrades under deployment timing conditions. Recommended fix: train with speed randomization (jitter/delay/spike augmentation).

1. Robustness Test — Timing Perturbations

Wraps the environment with timing perturbations. The agent runs normally — no internal intervention. Deployment scenarios (jitter, delay, spike) model realistic conditions. Stress scenarios (5x speed) test extreme resilience.

Robustness Under Timing Perturbations

Robustness Detail

Category	Scenario	Return (% nominal)	95% CI	RMSE ratio	Return Change
Deployment	Speed jitter (2 +/- 1)	28%	25%–32% ***	2.37x	+71.9%
Deployment	Observation delay (1 step)	2%	0%–4% ***	3.92x	+98.1%
Deployment	Mid-episode spike (1-5-1)	100%	91%–113%	1.08x	+0.2%
Stress	5x Speed (unseen frequency)	-12%	-14%–-10% ***	4.90x	+111.9%

Recommendation

Agent degrades under deployment timing conditions. Recommended fix: train with speed randomization (jitter/delay/spike augmentation).

Speeds tested: [1, 2, 3, 5, 8] | Episodes per condition: 30 | Intervention support: False

Generated by deltatau-audit v0.3.4