GenerativeDynamics
Docs/Reference

Planners and algorithm sources

Choose a task configuration, then change its planner while keeping the task, constraints, and evaluation tools. Ours identifies algorithms developed by the GenerativeDynamics authors. The list below follows the algorithm names in their papers; implementation keys are documented separately.

Algorithm What it does Source
MDOC · Ours Model-based diffusion with control-barrier-function projections inside dynamics rollouts. Paper · Original multi-robot project
MD-COAS · Ours Combines an augmented-Lagrangian feasibility prior, CFS projection, and adaptive constraint scheduling. Paper
2GO · Ours Shapes generative trajectory updates and exploration with active constraint geometry for constrained locomotion. Paper
MGA · Ours Combines RL sequence proposals, model-based evaluation, and realization-aware geometry for motion–impedance control. Implementation
MBD Uses known dynamics and Monte Carlo score estimates to optimize trajectories without demonstrations. Paper · Project
EB-MBD Introduces emerging barriers during model-based diffusion to retain useful samples under constraints. Paper
MPPI Updates sampled control sequences using exponential rollout-cost weights. Paper
DIAL-MPC Uses diffusion-inspired annealing to refine sampled controls in receding-horizon MPC. Paper
ATACOM Learns and executes actions in the tangent space of an augmented constraint manifold. Paper
ISSA Filters proposed actions through a derivative-free implicit-safe-set search. Paper
PegasusFlow Uses rolling denoising and weighted spline-basis updates for model-based trajectory optimization. Paper
DPCC Adds model-based projections and constraint tightening to learned trajectory diffusion. Paper
SafeDiffuser Applies control-barrier-function corrections within learned diffusion denoising. Paper

Pick a configuration

Algorithm Example configuration Compute path
MDOC configs/single_2d/mdoc.yaml JAX
MD-COAS configs/single_2d/mdcoas.yaml JAX + optimization dependencies
2GO configs/humanoid/corridor_2d/main/twogo_zone_a.yaml JAX + optimization dependencies
MGA configs/arm/peg_insert/main/mga.yaml JAX + simulation; learned prior assets depend on the recipe
MBD configs/single_2d/mbd.yaml JAX
EB-MBD configs/single_2d/ebmbd.yaml JAX
MPPI configs/humanoid/corridor_2d/main/mppi_zone_a.yaml JAX
DIAL-MPC configs/arm/peg_insert/baseline/dial.yaml JAX + simulation
ATACOM configs/arm/peg_insert/baseline/atacom.yaml JAX + simulation; trained tangent-action policy
ISSA configs/arm/peg_insert/baseline/issa.yaml JAX + simulation; trained nominal policy
PegasusFlow configs/arm/peg_insert/baseline/pegasusflow.yaml JAX + simulation
DPCC configs/d3il_avoiding/dpcc.yaml Optional Torch + D3IL; trained diffusion checkpoint
SafeDiffuser configs/d3il_avoiding/safediffuser.yaml Optional Torch + D3IL; trained diffusion checkpoint

See installation for extras and compatibility for the supported device paths. A paper link identifies the source algorithm; it does not imply that every upstream experiment, robot, or theoretical assumption is reproduced by this integration.

Names used in code

  • MD-COAS uses the historical cfsmbd / cfsmbd_full implementation keys. CFS-MBD is not a separate paper algorithm in this catalogue. The main mdcoas.yaml recipe enables adaptive scheduling; individual CFS configurations can select a different constraint projection or schedule.
  • 2GO uses the solver key 2go; its task configuration filenames use twogo. These refer to the same algorithm.
  • DIAL-MPC uses the solver key dial.
  • DPCC trajectory-selection settings, task dimensions, and constraint variants configure one integration rather than introducing new planners.
  • d3il_unified is an environment/solver execution adapter. It selects the algorithm through method_params.solver.
  • model_based_only, standalone_rl, and the MGA ablation configurations are comparison settings. They are not additional named paper algorithms.
  • CEM and CMA-ES implementations exist in the source tree, but this V1 catalogue lists algorithms with published task configurations. Source availability alone does not establish a supported task recipe.

Integration details

MDOC's original paper and project include multi-robot coordination through Conflict-Based Search. The V1 recipes in this repository expose its single-robot planning component; use the linked original project for the published multi-robot application.

The contact-control DIAL-MPC and PegasusFlow baselines retain their sampling mechanisms under the shared task and rollout interface. ATACOM trains and executes a policy in tangent-action coordinates. ISSA filters nominal policy actions through simulator queries. These are task integrations, so comparisons should report their configuration, learned assets, and total rollout cost.

SafeDiffuser uses a local CBF safety correction inside the vendored learned diffusion architecture. DPCC likewise uses its dedicated learned-model integration. Both require their optional Torch dependencies and trained assets.

MGA stands for Manifold Generative Annealing. Its manuscript is titled Manifold Generative Annealing for Contact Constraint-aware Impedance Control with Reinforcement Learning Prior. A public paper URL has not been verified, so this catalogue links the implementation. V1 qualifies its CPU path; GPU qualification and humanoid push/walking remain deferred.

Based on the framework documentation · docs/reference/planners.md · main · View source