Adversarial red‑team/blue‑team framework for AI2‑THOR embodied agents
The repo provides a full pipeline to generate, execute, and judge adversarial interactions between a red (attacker) LLM and a blue (defender) embodied robot in AI2‑THOR. It includes attack strategies, fine‑tuned attack models, simulation configs, and metrics export, enabling researchers to measure attack success rates across models. Useful for AI safety researchers and developers of embodied agents needing systematic robustness testing. Compared to ad‑hoc scripts, it offers an integrated, extensible benchmarking suite.
View on GitHub →Bharth2003/adversarial-embodied-ai