Choose a method

Begin with a task and the behavior you need. The four research methods share the framework, but they address different stages of constrained generative planning and control.

Your goal	Start with	Practical entry
Embed a complete planning loop in Python	2GO	Direct Python example, CPU with the optimization extra.
Learn configured runs and inspect artifacts	MBD	Planar CPU example, core package only.
Put safety projections inside model-based diffusion rollouts	MDOC	`configs/single_2d/mdoc.yaml`. The original project's multi-robot CBS application is a separate entry.
Combine feasibility scoring, CFS projection, and adaptive scheduling	MD-COAS	`configs/single_2d/mdcoas.yaml`; then D3IL avoidance with its additional setup.
Shape updates and exploration using active constraint geometry	2GO	Humanoid corridor or quadruped stepping stones.
Refine policy proposals for contact-rich motion–impedance control	MGA	Peg insertion or surface scanning; install simulation dependencies and obtain the recipe's frozen learned assets.

Keep paper names and implementation names distinct

MDOC uses mdoc.
MD-COAS uses the historical cfsmbd / cfsmbd_full implementation family. The main mdcoas.yaml selects adaptive scheduling. MD-COAS-F and MD-COAS-A are configuration variants, not additional framework packages.
2GO uses the method key 2go; configuration filenames also use twogo.
MGA uses mga. Learned priors and reliability settings are task-specific; read the selected configuration and asset contract.

Switch a planner without changing the question

Use configurations from the same task family. Keep the task, condition, seeds, metric definitions, and compute protocol fixed while comparing methods. Do not assume changing the method field alone is sufficient: method-specific parameters, constraint filters, or learned checkpoints may differ. The planner catalogue links the supported configurations and comparison methods.

Check the supported path

V1 is JAX-first. CPU is the portable entry; CUDA qualification varies by task. DPCC and SafeDiffuser use specialized optional Torch integrations. The broader research gallery includes hardware and humanoid contact demonstrations beyond the qualified V1 recipe set. Consult compatibility and release scope before selecting a device or robot.

Sources

Adapted from the repository README, planner catalogue, recipes, and release scope.

Based on the framework documentation · Website guide · main