Shared interfaces.
Room for the next method.
GenerativeDynamics brings learning, generative inference, and control together through explicit task, solver, constraint, and execution boundaries.
Dynamics · Objectives · Geometry
Robot models, task costs, collision geometry, and physical constraints.
Planners · Priors · Constraint solvers
Interchangeable algorithms, learned proposals, and geometry-aware updates.
Controllers · Governors · Observers
Replanning, feedback, safety components, and inspectable results.
Compose around
your robot task.
Planning
Use a solver directly from Python or select one through a configuration. Share dynamics, objectives, geometry, and evaluation across comparisons.
Control
Connect plans to receding-horizon execution, controllers, governors, action filters, recovery components, and robot I/O.
Learning
Train PPO/SAC policies, reuse policy checkpoints as MGA horizon priors, or work with learned trajectory diffusion through optional integrations.
Many algorithms.
A common place to work.
Research methods, model-based planners, sampling control, and learned integrations share the surrounding task and experiment tooling.
| Method | Family | Role in the stack |
|---|---|---|
| MDOC | Our research | Model-based diffusion with CBF projections inside dynamics rollouts. |
| MD-COAS | Our research | iALM soft feasibility, CFS correction, and adaptive constraint scheduling. |
| 2GO | Our research | Active geometry shapes refinement and reverse stochastic exploration. |
| MGA | Our research | Learned proposals and feedback-conditioned motion–impedance refinement. |
| MBD | Model-based generation | Known dynamics and Monte Carlo score estimates, without demonstrations. |
| EB-MBD | Model-based generation | Emerging barriers retain informative candidates under constraints. |
| MPPI | Sampling MPC | Cost-weighted control samples for receding-horizon optimization. |
| DIAL-MPC | Model-based generation | Diffusion-inspired annealing for online control-sequence refinement. |
| ATACOM | Contact / constraints | Tangent-space actions on an augmented constraint manifold. |
| ISSA | Contact / constraints | Derivative-free action filtering through implicit safe-set search. |
| PegasusFlow | Model-based generation | Rolling denoising with weighted spline-basis trajectory updates. |
| DPCC | Learned diffusion | Learned trajectory diffusion with model-based projection and tightening. |
| SafeDiffuser | Learned diffusion | CBF corrections during learned trajectory denoising. |
MD-COAS uses the historical CFS-MBD implementation. 2go and twogo are aliases. Dependencies, devices, and checkpoints vary by method; consult the compatibility and planner references.
Choose your starting point.
Planar navigation
CPU-first constrained trajectory planning.
7-DoF avoidance
Arm motion through constrained workspaces.
Quadruped stepping stones
Discrete footholds and governed execution.
Humanoid corridor
Body motion, posture, and arm clearance.
Peg insertion
Contact transitions and recovery under variation.
Surface scanning
Surface tracking, compliance, and force regulation.