Поиск исследовательских идей для робототехники и воплощённого ИИ

★ 7.1 · research

idea-discovery-robot is a Claude Code skill that runs a full idea-discovery pipeline for robotics and embodied AI research. It chains four sub-skills in sequence: a literature survey focused on CoRL, RSS, ICRA, IROS, and RA-L; robotics-aware idea generation that factors in embodiment type, task family, and observation/action interface; novelty verification; and critical review via an external model through Codex MCP. Every phase defaults to simulation-first execution — up to three pilot ideas are validated in simulators or on offline logs, and real-robot runs require explicit approval. The skill builds a structured landscape matrix across benchmarks such as ManiSkill, RLBench, Isaac Lab, and CALVIN, and explicitly surfaces recurring failure modes, saturated benchmarks, and hidden infrastructure dependencies. It is aimed at researchers working on manipulation, locomotion, navigation, drone, or humanoid learning who need to move from a broad direction to a falsifiable, benchmark-grounded research question.