Simulusion
An MIT-licensed open-source PyTorch world-model library that integrates world models such as Dreamer, IRIS and DIAMOND through a unified WorldModel protocol, providing latent-space probes, environment adapters and a training toolchain.
Timeline (last 90 days)
Simulusion is an MIT-licensed PyTorch open-source world model library aimed at encapsulating different world models with a unified WorldModel protocol.
Simulusion's WorldModel protocol consists of three methods: encode(frames, actions), step(state, action) and rollout(state, actions).
Simulusion has integrated the RSSM family (Dreamer, PlaNet) and the token Transformer family (IRIS), and diffusion models (DIAMOND) are also within its support scope.
Simulusion includes three representation quality measurement tools: linear probes, DCI and MIG.
Simulusion's environment adaptation layer uniformly supports DeepMind Control, Atari, Gymnasium, MuJoCo and Brax through a single make_env entry point.
The Simulusion protocol includes a decode step, which poses a compatibility challenge for TD-MPC-style models that do not need decoding.
The author posted in the MBRL community on Reddit to solicit feedback on interface generality and roadmap priorities.
A world model can be trained separately from the agent using only recorded episode data, and after training any agent that follows the protocol can directly use it for planning.
Simulusion's project address is https://github.com/paramthakkar123/simulusion, and its website is https://simulusion.vercel.app/.
All Facts (9)
Simulusion's project address is https://github.com/paramthakkar123/simulusion, and its website is https://simulusion.vercel.app/.
50%UnverifiedSimulusion's WorldModel protocol consists of three methods: encode(frames, actions), step(state, action) and rollout(state, actions).
50%UnverifiedSimulusion has integrated the RSSM family (Dreamer, PlaNet) and the token Transformer family (IRIS), and diffusion models (DIAMOND) are also within its support scope.
50%UnverifiedSimulusion includes three representation quality measurement tools: linear probes, DCI and MIG.
50%UnverifiedSimulusion's environment adaptation layer uniformly supports DeepMind Control, Atari, Gymnasium, MuJoCo and Brax through a single make_env entry point.
50%UnverifiedThe Simulusion protocol includes a decode step, which poses a compatibility challenge for TD-MPC-style models that do not need decoding.
50%UnverifiedSimulusion is an MIT-licensed PyTorch open-source world model library aimed at encapsulating different world models with a unified WorldModel protocol.
50%UnverifiedA world model can be trained separately from the agent using only recorded episode data, and after training any agent that follows the protocol can directly use it for planning.
50%UnverifiedThe author posted in the MBRL community on Reddit to solicit feedback on interface generality and roadmap priorities.
50%