Users are excited that GPT can build a MuJoCo-like physics engine from scratch in under 10 prompts, citing impressive agent training with PPO and the Python DSL's quality-of-life gains over XML.
Based on 3 visible X reactions from 3 accounts; directional sample.
Ask a question below.
Published answers will appear here.
@cloneofsimo the python DSL instead of xml is the kind of quality of life improvement that actually makes me want to build environments. xml for physics defs always felt like unnecessary friction
+ you can train control-following agent with PD control via PPO!! https://x.com/cloneofsimo/status/2077029757158592717/video/1
@cloneofsimo Would be extremely impressive if we can implement it with warp
GPT 5.6 Sol made complete mujoco-like physics engine from scratch: gym-like api, c++ implemented, native headless render, web render, python native DSL to define environments instead of annoying XML ngl im not claiming its better than mujoco or its completely original, but it is pretty insane now you have some min-physics-engine implementation that is *actually legit??* because its based on wasm very easy cross platforming. EVERYTHING was made with < 10 prompts https://cloneofsimo.github.io/lavendersim
GPT 5.6 Sol made mujoco-like physics engine from *completely* scratch: gym-like api, c++ implemented, native headless render, web render, python native DSL to define environments instead of annoying XML ngl im not claiming its better than mujoco or its completely original, but it is pretty insane now you have some min-physics-engine implementation that is *actually legit??* EVERYTHING was made with < 10 prompts https://cloneofsimo.github.io/lavendersim
@cloneofsimo the python DSL instead of xml is the kind of quality of life improvement that actually makes me want to build environments. xml for physics defs always felt like unnecessary friction
+ you can train control-following agent with PD control via PPO!! https://x.com/cloneofsimo/status/2077029757158592717/video/1
@cloneofsimo Would be extremely impressive if we can implement it with warp
GPT 5.6 Sol made complete mujoco-like physics engine from scratch: gym-like api, c++ implemented, native headless render, web render, python native DSL to define environments instead of annoying XML ngl im not claiming its better than mujoco or its completely original, but it is pretty insane now you have some min-physics-engine implementation that is *actually legit??* because its based on wasm very easy cross platforming. EVERYTHING was made with < 10 prompts https://cloneofsimo.github.io/lavendersim
GPT 5.6 Sol made mujoco-like physics engine from *completely* scratch: gym-like api, c++ implemented, native headless render, web render, python native DSL to define environments instead of annoying XML ngl im not claiming its better than mujoco or its completely original, but it is pretty insane now you have some min-physics-engine implementation that is *actually legit??* EVERYTHING was made with < 10 prompts https://cloneofsimo.github.io/lavendersim
Users are excited that GPT can build a MuJoCo-like physics engine from scratch in under 10 prompts, citing impressive agent training with PPO and the Python DSL's quality-of-life gains over XML.
Based on 3 visible X reactions from 3 accounts; directional sample.
Ask a question below.
Published answers will appear here.