Choose safer futures with compositional physics.

Compositional Port-Hamiltonian World Models for Structured Dynamics Transfer

One observed state. Multiple possible futures. Typed physics tells you what each future costs.

C-PHAST composes reusable port-Hamiltonian blocks into a deployed world model. It rolls out candidate futures and exposes energy, support, stability, and residual readouts for downstream planners.

energy contact/support stability residuals
Watch demos Paper How it works Citation
Same observed Go2 state. Different supplied futures. C-PHAST scores the predicted physical consequences before simulator-level failure.

One state, three futures

The output is a choice, not just a trajectory.

The same learned typed block library can be queried under different candidate actions or contact patterns. Each rollout returns a physical commitment profile.

Filmstrip of C-PHAST candidate futures from the same Go2 state.
Same observed state, three supplied futures: continue, recover with support, or fail. The cards below summarize what the typed readouts make visible.
1

Fast but risky

Low energy and low support, but near the stability boundary.

2

Robust but costly

More contact and effort, but a much larger safety margin.

3

Unsafe

Residuals and stability channels predict contact failure.

Playable evidence

Watch the rollouts.

Four distinct checks: futures, safety, transfer, and baseline failure.

Candidate futures. Same observed state, different physical commitments.
Push boundary. The same policy recovers, struggles, or fails as perturbation grows.
Embodiment transfer. Source ANYmal-D and target Go2 side-by-side.
Baseline failure. Target rollouts reveal drift and divergence.

How it works

Reusable typed physics, deployment-specific composition.

C-PHAST declares the deployment, infers typed phase state from history, reuses a shared PHASTCore library, composes under the target interconnect, and returns native typed readouts for decision layers.

C-PHAST rollout contract from DeploymentSpec to typed readouts and planner logic.
Typed local dynamics are reused; only the deployed composition changes.

Same principle beyond robots

Microgrid redeployment is the electrical analogue.

A microgrid also changes through topology, sensors, and active assets. C-PHAST reuses typed pH asset blocks and recomposes them under a new target graph. The readouts become voltage margin, current loading, loss, and model-fit residuals.

Microgrid redeployment figure showing candidate network futures from the same pH asset library.
Candidate network futures from the same typed pH asset library.

Citation

BibTeX

@article{bhardwaj2026cphast,
  title   = {Compositional Port-Hamiltonian World Models for Structured Dynamics Transfer},
  author  = {Bhardwaj, Shubham and Bajaj, Chandrajit},
  journal = {arXiv preprint},
  year    = {2026}
}