Compositional port-Hamiltonian world models

Compare physical futures before choosing an action.

C-PHAST learns reusable physics blocks, recomposes them for a target deployment, and turns candidate rollouts into named physical consequences.

The model does not emit one opaque score. It predicts quantities a planner can reason about: energy use, support/contact margin, sensing value, voltage or current margin, and residual mismatch.

Real robot result Interactive explorer Action-card interface Paper Citation
Same observed Go2 state, different candidate futures. The interface makes the physical tradeoff visible before a downstream planner chooses.
0.880 / 0.860 Pearson / Spearman agreement on held-out real Spot windows
ANYmal-D to Go2 cross-robot structured dynamics transfer in Isaac Lab
FHN + microgrid repeated-cell and typed node-edge recomposition beyond legged robots

What the model exposes

From observables to action cards.

A typed consequence is a named physical quantity with a role: reward, cost, safety envelope, or residual check. Objective cards reuse the same predicted consequences, then change weights or veto rules.

Pipeline from raw observables to typed consequences, planner factors, and action cards.
The same candidate forecast can support efficient, safe, mapping, balanced, or inspection objectives without retraining the dynamics model.
1

Observe

Recent robot, circuit, or grid measurements refresh the deployed state.

2

Roll out

Candidate futures are simulated with the target blocks, graph, ports, and sensing interface.

3

Name the consequences

Rollouts become physical readouts such as work, support margin, map value, voltage margin, or residual mismatch.

4

Choose by objective

Downstream objectives turn the same consequence vector into action cards, rankings, or vetoes.

Real robot evidence

Spot-with-arm logs test the telemetry-to-card path.

Six teleoperated Spot-with-arm episodes are replayed offline. C-PHAST reads recent telemetry and synchronized perception summaries, predicts named future consequences for a possible next behavior, and writes objective-conditioned action cards beside the logged run.

This is shadow mode: no autonomous commands are sent to Spot. The result validates the real sensing and decision interface, while the Isaac experiments below validate branchable same-state alternatives in simulation.

1,070 five-second behavior windows
173 held-out windows
5 cards efficient, safe, mapping, balanced, inspection
Real Spot-with-arm shadow-mode query from logged context to candidate behavior and action-card selection.
Real Spot query: logged context, candidate behavior descriptors, and objective-specific selected futures.

Branchable rollout evidence

Simulation asks the counterfactual question.

Real logs show whether the interface works on hardware data. Simulator rollouts let us hold the state fixed and ask what would happen under alternative candidate futures.

Candidate futures. Same state, different commanded futures, different physical consequences.
Push boundary. The same policy recovers, struggles, or fails as perturbation grows.
Embodiment transfer. Train on ANYmal-D, deploy on Unitree Go2 without retraining.
Baseline contrast. Unstructured target rollouts drift where the typed pH composition stays physically anchored.

Same contract, different physics

Transfer is expressed in each domain's own language.

C-PHAST is not a single robot trick. The same contract appears as repeated joint or leg blocks, repeated excitable cells, and typed converter-line blocks in a power network.

ANYmal-D to Go2 transfer snapshot.

Legged robots

Reusable mechanical blocks are recomposed for a target embodiment and support/contact pattern.

C-PHAST deployment contract figure.

Repeated excitable cells

FitzHugh-Nagumo chains test whether repeated cell blocks transfer under changes in cell count and diffusive graph.

Microgrid redeployment visual with candidate network futures.

DC microgrids

DGU node blocks and line edge blocks are recomposed under topology and sensing changes.

C-PHAST rollout contract from deployment specification to typed readouts and planner logic.
The learned block library stays fixed; the deployment specification declares the target blocks, interconnection, observation interface, and readout roles.

Model contract

Reusable blocks, deployment-specific wiring.

The method separates local physics from the way a target system is assembled. This is why the paper can compare subsystem-count changes, cross-robot transfer, repeated excitable circuits, and microgrid topology or sensing changes under one interface.

  • Local blocks store and dissipate energy.
  • Ports describe how power enters, leaves, or moves between blocks.
  • Interconnections wire the deployed system without changing the learned local library.
  • Typed readouts expose what a planner or operator needs to compare futures.

Citation

BibTeX

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