📡 Idea Radar 每天 8:00 / 18:00 推送 · 18:00 生成精选
reddit · SideProject · 54 分 ← 返回列表

I built a "compiler" for robot simulation worlds, because LLM-generated were not good enough

I work on vision and RL for robotics, and I can't crash a drone or robot in real life every time a policy fails. So simulation is the safe place to test, but I'm not a 3D artist and it took me a long time to build something from scratch in IsaacSim I asked Claude Code to build the worlds in it but there were some problems like: assets clipped through walls and robots spawned under the map. So I spent some time on building a "compiler" for physical world: the agent writes

五维评分

付费意愿
45
需求真实度
55
个人可落地
60
竞争空间
50
新鲜度
75

AI 分析

痛点:机器人/RL团队需要快速生成可物理验证的仿真世界,但手工搭建 Isaac Sim 场景耗时,且 LLM 直接生成常出现穿模、机器人生成到地图下等错误。

目标用户:机器人视觉/强化学习研究者、仿真工程师、机器人初创团队

证据:I can't crash a drone or robot in real life every time a policy fails. So simulation is the safe place to test, but I'm not a 3D artist and it took me a long time to build something from scratch in IsaacSim ... assets clipped through walls and robots spawned under the map.

变现:按生成场景/API 调用收费的 SaaS,企业私有化部署与领域模板订阅,或提供仿真资产与验证服务。

MVP:2周内做仓库/室内等单一场景的 prompt→规格 DSL→坐标求解→几何/物理校验→导出 USD/Isaac Sim 的流水线,先服务 5 个种子机器人团队。

竞品:NVIDIA Isaac Sim、Omniverse、Gazebo、Unity Robotics、Unreal Engine

摘要:作者为机器人视觉/RL 开发者,因真实测试危险且手工搭 Isaac Sim 场景耗时,做了把自然语言/声明式规格编译成仿真世界的工具。LLM 直接生成会出现资产穿墙、机器人出生在地图下等问题,他的方案加入求解与几何/物理校验。帖子未给出付费或收入证据,需求真实但市场验证仍弱。