Generate adversarial scenarios to stress-test physical AI policies.
npx skills add https://github.com/nebius/nebius-physical-ai --skill scenario-gen-nebius
Please help me install this Agent Skill. Skill: scenario-gen Source: https://github.com/nebius/nebius-physical-ai/tree/main/skills/tools/scenario-gen Command: npx skills add https://github.com/nebius/nebius-physical-ai --skill scenario-gen-nebius
This skill addresses the difficulty of identifying edge-case failures in physical AI policies by automating the generation of adversarial scenarios that maximize policy violation rates.
Use the scenario-gen skill to generate an adversarial set for the policy located at the specified S3 URI and save the results to the output path.
High-intent search queries and answers about installing and using this skill.
Adversarial scenario generation uses RL-based or heuristic-based agents to perturb environments and surface hard-case failures for regression testing. This skill automates that process, maximizing policy violation rates to identify edge-case failures.
Automated severity ranking scores generated adversarial scenarios based on failure severity and diversity. This scoring mechanism prioritizes hardening efforts by highlighting the most critical policy violations discovered during simulation.
Yes, Isaac Lab RL is supported as a pluggable backend for training adversarial agents. The skill orchestrates end-to-end workflows to stress-test and harden physical AI policies against navigation and other failure modes.
Visualizing physical AI policy failures is handled through Rerun integration, which provides visual heatmaps and time-series analysis. These artifacts help analyze how policies break down under adversarial scenario conditions.
Yes, the skill supports heuristic search as a pluggable backend for deterministic environments. This allows you to generate adversarial scenarios without requiring reinforcement learning, ensuring deterministic reproducibility.
You generate an adversarial set for your target policy, rank the scenarios by failure severity, and save the results to an output path. These ranked failure scenarios can then be fed directly into a retraining loop.
robotics-playgroundGuides RL locomotion policy development for Unitree G1 using NVIDIA Isaac Sim and Isaac Lab.
cxnaiveOptimize language models with SFT, DPO, RLOO, and GRPO post-training alignment.
uw-ssecRenders large-scale data visualizations using Datashader, HoloViews, and Panel.
robotlearning123Train and evaluate reinforcement learning agents with PyTorch and Gymnasium.
