Robust Safe Algorithms¶

Robust safe methods add a cost constraint and make it hold under shifted dynamics. They are implemented in the online setting on a shared PPO learner and run on the Isaac Lab tasks.

Features¶

Feature	Robust Safe Algorithms
Interface	The Isaac Lab training and evaluation recipe
Shared learner	One PPO implementation for every method
Constraint	A cost budget on joint-limit violations
Tasks	Unitree G1 locomotion and Franka drawer manipulation

Supported methods¶

Method	Family	Base	Claimed robustness	Shifted-env rollout	Adversarial network	Learned model
RAMU	Robust safe	PPO	Dynamic
SPiDR	Robust safe	PPO	Dynamic	✓		✓

Claimed robustness is the shift the original paper targets.

Run a method¶

The methods of this group are trained and evaluated with the Isaac Lab recipe, on one PPO implementation that every method shares.

Method pages¶

RAMU: A risk measure over sampled next-state perturbations, applied to reward and cost targets.
SPiDR: Constrained learning under domain randomisation with a pessimistic cost penalty.