CaLiSym: Learning Symplectic Dynamics of Real-World Systems through Structured Canonical Lifts

Aristotelis Papatheodorou; Gerard J. Milburn; Ioannis Havoutis; Natalia Ares; Pranav Vaidhyanathan

Symplectic learning can be extended to real robots by lifting states and energy ports into a structured canonical phase space where the map stays exactly symplectic.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-10 20:32 UTC pith:GIQOX5KX

load-bearing objection Solid engineering paper: structured lift + exact SympNet works on real open robots and beats larger baselines; the contact-rich geometric story is weaker than the abstract claims, but the empirical result still stands. the 3 major comments →

arxiv 2607.06824 v2 pith:GIQOX5KX submitted 2026-07-07 cs.RO cs.LG

CaLiSym: Learning Symplectic Dynamics of Real-World Systems through Structured Canonical Lifts

Aristotelis Papatheodorou , Pranav Vaidhyanathan , Natalia Ares , Ioannis Havoutis , Gerard J. Milburn This is my paper

classification cs.RO cs.LG

keywords robot dynamics learningsymplectic neural networksgeometric machine learningstructure-preserving learningHamiltonian systemscontact-rich roboticscanonical liftsmodel-based control

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Exact symplectic predictors have been reliable for closed conservative systems, but real robots continually exchange energy through actuation, damping, and contact, so the measured state is not a closed Hamiltonian flow. CaLiSym argues that the fix is not to drop the geometric prior, but to move it: embed the physical state and its ports into a gauge-fixed section of a larger canonical phase space, evolve there with an exactly symplectic map, then project back. The lift is algebraic and explicit, so inference stays a single forward pass with no latent history, decoder, or ODE solve. Instantiated with compact SympNet-style layers (including a B-spline ridge variant for low-dimensional lifts), the method improves out-of-distribution autoregressive rollout on a forced damped double pendulum, a real quadrotor, and a contact-rich quadruped while using fewer parameters than black-box and physics-informed baselines, and while preserving the lifted symplectic form to numerical precision. A sympathetic reader cares because this would make geometry-preserving world models usable inside closed-loop planners on the open, non-conservative systems that robots actually are.

Core claim

The paper claims that symplectic structure need not be imposed on the measured physical state. By embedding state and physical ports into a structured lifted canonical phase space and learning an exactly symplectic discrete map there, then projecting and re-embedding with the next port, one obtains geometry-preserving predictors for controlled, dissipative, and contact-rich robot dynamics. Empirically this yields lower out-of-distribution autoregressive error with smaller models, while the lifted map preserves the symplectic form to machine precision.

What carries the argument

The structured canonical lift with gauge-fixed rollout: at each step the physical state and ports are embedded into a data section of the lifted phase space (auxiliary fibers zeroed, ports fixed), an exactly symplectic map advances the lifted state, the physical coordinates are projected out, and the next ports re-embed the prediction. Symplecticity is enforced only in the lift; the induced physical map may be forced or dissipative.

Load-bearing premise

The one-step physical transition must admit a smooth symplectic realization on that lifted data section, and the ports used for re-embedding (including contact forces) must be available at every step; if either fails, the geometric story no longer justifies the procedure.

What would settle it

Train the same lifted symplectic predictor and matched baselines on a system with hard unmodeled impacts or missing port measurements, then check whether out-of-distribution multi-horizon autoregressive error still improves while the lifted Jacobian residual stays near machine precision; collapse of either metric under those conditions would refute the central claim.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

Symplectic inductive bias is no longer limited to closed conservative mechanical systems in robotics learning.
Markovian single-step maps with exact lifted geometry can outperform larger context-window sequence models on OOD long-horizon rollout.
Ports treated as canonical variables make energy exchange between robot and environment an explicit bookkeeping object for passivity-style control.
The same lift–evolve–project–re-embed procedure can transfer across dissipative, aerial, and contact-rich platforms without changing the geometric prior.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

If ports are only partially observed, the construction suggests learning a lightweight port estimator that still feeds a symplectic lift rather than abandoning the geometry.
The energy-exchange bookkeeping exposed by the ports is a natural hook for certificate-based or energy-shaping controllers that currently rely on hand-derived Hamiltonians.
Hard hybrid contact mode switches may require piecewise or switched lifts rather than a single smooth section, a concrete next stress test of the method.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Desk Editor's Note

Solid engineering paper: structured lift + exact SympNet works on real open robots and beats larger baselines; the contact-rich geometric story is weaker than the abstract claims, but the empirical result still stands.

read the letter

The useful takeaway is simple. They keep an exact symplectic discrete map by lifting the measured state and ports into a gauge-fixed canonical section, evolve with GR/GRB-SympNet, then project and re-embed. That is the actual novelty relative to SympNets, HNNs, and port-Hamiltonian learning. It is explicit, Markovian, and cheap at inference.

What they do well: architecture is symplectic by construction (residuals ~1e-15–1e-16), training is clean teacher-forcing, and the OOD protocol is better than the usual partial-trajectory split. Double pendulum (energy-regime OOD), real quadrotor flights, and GRANDTOUR quadruped all show lower AR MSE with fewer parameters than MLP, Transformer, RWM-TF, DHNN, and D-SymODEN. GRB-SympNet is a sensible low-dim spline ridge variant; the universal-approx appendix is standard but correctly scoped to the lifted map. Citations are fair.

Soft spots, in proportion. The stress-test is right on the geometry: Appendix C needs F(x,e)=(f_e(x),e) to be a local diffeomorphism so the cotangent lift recovers the physical map. Hard impacts make the one-step map discontinuous in momentum, so the quadruped is not literally a cotangent lift of hybrid dynamics. It is an algebraic embedding of measured contact forces plus a symplectic network that fits well. Ports (including contacts) must be supplied at every re-embed step, so this is not a pure state predictor. No error bars, no code yet. Those are real limits; they do not erase the OOD gains or the architectural guarantee on the lifted map.

Who it is for: people building learned models for MPC / model-based RL on forced or contact robots who want geometric inductive bias without BPTT or latent history. Worth a serious referee. I would engage, cite the lift construction, and flag the hybrid-geometry caveat when using the quadruped claim.

Referee Report

3 major / 5 minor

Summary. CaLiSym proposes to extend exact symplectic dynamics learning beyond closed conservative systems by embedding the measured physical state and interaction ports (actuation, contacts) into a gauge-fixed section of a structured lifted canonical phase space, evolving that state with an exactly symplectic discrete map (GR-SympNet or the new GRB-SympNet), then projecting and re-embedding. The lift is algebraic and Markovian; inference is a single forward pass. The paper reports OOD autoregressive gains and parameter efficiency on a controlled dissipative double pendulum, a real quadrotor, and a contact-rich real quadruped, with lifted Jacobian residuals at numerical precision, plus a universal-approximation result for full-ridge GRB-SympNet.

Significance. If the construction is sound, this is a practically important step: it moves the strongest geometric prior in dynamics learning (exact symplecticity) into the regime of forced, dissipative, and contact-constrained robots that dominate model-based control. Strengths that should be credited include (i) architectural rather than soft-loss enforcement of the symplectic form, with residuals reported at 1e-15–1e-16; (ii) proper OOD splits by energy regime, held-out flights, and unseen missions rather than temporal holdouts of the same trajectories; (iii) teacher-forcing training with fully Markovian AR evaluation; (iv) an explicit multi-horizon ID/OOD table; and (v) a stated approximation theorem for GRB-SympNet. The framework is also architecture-agnostic in principle and parameter-efficient relative to the reported baselines.

major comments (3)

[Appendix C; §IV.B–D; §V.D] Appendix C (and the modeling justification in §IV.B–D) rests on a sufficient local condition that the joint map F(x,e)=(f_e(x),e) is a local diffeomorphism so that a cotangent lift recovers the physical transition after projection onto the gauge-fixed section. For the quadruped, hard impacts make the one-step physical map discontinuous in momentum; F is then not continuous, let alone a diffeomorphism. In that regime Eq. (24) and the exact Hamiltonian flow of GR-SympNet no longer have a rigorous interpretation as a cotangent lift of the observed hybrid dynamics—the procedure becomes an algebraic port embedding plus a symplectic network that fits data. The double-pendulum and quadrotor cases are smoother and better covered. The paper’s strongest claim (symplectic learning for contact-rich real robots) therefore needs either a hybrid-systems justification, a clearly weakened geometric claim
[Algorithm 1; Eq. (63); §V.D; Introduction] Algorithm 1 re-embeds with the next port e_{k+1} at every step. For the quadruped this includes contact forces f_c (Eq. 63). The model is therefore not a pure state predictor: long-horizon AR evaluation and any closed-loop use require contact ports to be measured, estimated, or treated as decision variables at every step. The introduction argues this matches MPC decision variables, which is fair for planning, but the OOD AR numbers in Fig. 4 and the abstract’s “dynamics prediction” framing should state this dependency explicitly and discuss failure modes when contact ports are noisy or missing. Without that, the contact-rich result overstates what is being predicted from state alone.
[§V; Table V; Eq. (56)] There is no ablation that isolates the structured lift from capacity and training choices: e.g., the same GR-SympNet/GRB-SympNet backbone with ports concatenated as ordinary inputs but without the gauge-fixed section, projection, and re-embedding, or a non-symplectic lifted baseline of matched size. Loss weights are also highly system-specific (Table V: w_μ=65, w_fc=55 on the quadruped vs. much lighter port weights elsewhere). Given that OOD AR MSE is the headline metric, at least one controlled ablation on the lift and a short sensitivity check on port/section weights would make the causal claim—that gains come from the geometric construction rather than port injection and hyperparameter tuning—much stronger.

minor comments (5)

[Figures 2–4] Fig. 2–4 bar plots report single OOD AR MSE values with no error bars, seeds, or variance across windows/missions. Even a short note on variability would help assess whether the 12.7% quadruped gain is stable.
[Appendix A, Table III] Table III shows OOD MSE growing to ~101 at 800 steps while ID stays small. The text correctly separates numerical stability from distribution-shift compounding; a brief comparison of the same multi-horizon protocol for the strongest baseline would make that separation more convincing.
[§IV.B; Eqs. (16)–(20), (61)–(63)] Notation for the lift is dense (Q,P vs. b,ζ; μ_u vs. τ; λ_c vs. f_c). A single small table mapping physical ports to lifted coordinates for each of the three systems would reduce reader load.
[Footnote †; Appendix D] The manuscript promises open-source release upon acceptance; if code and data splits can be linked in revision, that would strengthen reproducibility of the OOD partitions and symplectic residual checks.
[Headers; §I; References] Minor typos and formatting: “V AIDHY ANATHAN” spacing in running headers; “markovian” should be “Markovian”; arXiv IDs in the reference list for GRANDTOUR and related works should be double-checked for consistency with the cited versions.

Circularity Check

0 steps flagged

No significant circularity: symplecticity is architectural by construction, OOD gains are external empirical comparisons, and the lift is a self-contained geometric construction rather than a fit renamed as prediction.

full rationale

Walking the load-bearing chain: (i) the physical system is open (actuation/dissipation/contact), so the measured map on T*Q need not be symplectic; (ii) state and ports are algebraically embedded into a gauge-fixed section of a lifted canonical phase space; (iii) an exactly symplectic discrete map (GR-/GRB-SympNet layers as exact ridge Hamiltonian flows) advances the lifted state; (iv) projection and re-embedding recover the physical one-step predictor. Symplecticity of Φ is enforced by architecture (each layer is an exact Hamiltonian flow; composition preserves the form), and the reported Jacobian residual ϵ_JΩ at numerical precision is framed as verification of that construction, not as a data-driven discovery. Empirical claims (lowest OOD AR MSE on double pendulum, quadrotor, quadruped vs MLP/Transformer/RWM-TF/DHNN/D-SymODEN, with smaller parameter counts) are ordinary held-out comparisons on distinct energy/flight/mission regimes, not quantities forced by fitting a parameter then re-reporting it. Appendix C’s universal-approximation argument reduces GRB-SympNet density to the external Hénon/SympNet route under an explicit local-diffeomorphism assumption; that assumption may be false for hard impacts (a correctness risk), but it is not a circular reduction of a claimed prediction to its own inputs. Self-citation of MetaSym appears only as related work and is not used as a uniqueness theorem or load-bearing premise. Loss weights and ridge widths are free hyperparameters and do not make reported OOD MSE equal a fitted constant by construction. No self-definitional loop, fitted-input-as-prediction, uniqueness-from-authors, or ansatz-smuggled-via-self-citation steps are present.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 2 invented entities

The paper rests on standard symplectic geometry and cotangent-lift constructions from geometric mechanics, plus the modeling choice that ports are observed and that the physical map admits a lifted symplectic realization. Free parameters are ordinary training hyperparameters and architecture widths; invented entities are the CaLiSym procedure and GRB-SympNet, which are algorithmic rather than new physical objects.

free parameters (4)

loss weights (w_q, w_p, w_sec, w_mu, w_fc)
Hand-chosen per system (e.g. quadruped w_mu=65, w_fc=55 vs pendulum w_q=10); they shape which ports the fit emphasizes and therefore affect reported MSE.
ridge dimension m and layer counts L
Architecture capacity knobs (m=2 for GRB pendulum; widths 36/66 for GR) chosen for parameter efficiency; not derived from first principles.
B-spline degree, intervals, knot range for GRB-SympNet
Cubic splines, 12 intervals on [-3.5,3.5] set by hand for the double-pendulum experiment.
optimizer, learning rates, TF horizon, batch size
AdamW/Muon schedules and 100-step teacher-forcing horizon are experimental choices that influence final OOD numbers.

axioms (5)

standard math Exact Hamiltonian flows and their compositions preserve the canonical symplectic form
Used throughout §IV-E and Appendix B to claim numerical-precision symplecticity of GR/GRB-SympNet layers.
domain assumption Open robotic dynamics of form (12) admit a cotangent-lift Hamiltonian (24) on an augmented phase space whose projection recovers physical evolution
Core modeling step in §IV-B/C; justified by geometric-mechanics literature but not proved for every contact mode.
domain assumption Control and contact ports e_k are known at each discrete step for gauge-fixed re-embedding
Algorithm 1 and §IV-D require e_{k+1} to form Z_{k+1}; realistic for commanded torques, stronger for measured contact forces.
domain assumption Generalized momenta can be obtained from velocities via a known mass matrix M(q)
§IV-F; analytical or robot-model M(q) is assumed available for all three systems.
ad hoc to paper Teacher-forcing one-step training plus autoregressive evaluation fairly isolates architecture quality
§V-A deliberately avoids AR training; comparison validity rests on this protocol choice.

invented entities (2)

CaLiSym structured gauge-fixed canonical lift and lift-evolve-project-re-embed rollout no independent evidence
purpose: Move exact symplecticity off the physical state onto an augmented space so open systems can still use symplectic predictors
The specific section S_ek, embedding sigma_e, and re-embedding schedule are the paper's central construction; independent evidence is the empirical OOD gains and residual tables, not an external physical discovery.
GRB-SympNet (generalized-ridge B-spline SympNet) no independent evidence
purpose: Add local spline expressivity to exact ridge symplectic layers for low-dimensional lifted systems
New architecture variant; universality sketched in Appendix C under full-ridge and ridge-structured assumptions.

pith-pipeline@v1.1.0-grok45 · 31682 in / 3432 out tokens · 45617 ms · 2026-07-10T20:32:37.226513+00:00 · methodology

0 comments

read the original abstract

Physics-informed learning promises data-efficient and stable dynamics prediction, yet its strongest geometric guarantees have largely remained confined to closed conservative systems. This excludes robotic systems of interest, where actuation, dissipation, and constraints exchange energy and momentum with the environment. We introduce CaLiSym, a lightweight framework that extends symplectic learning to such systems by changing where the geometric prior is imposed. Rather than enforcing symplecticity on the measured state, CaLiSym embeds the state and its ports into a lifted phase space, where the dynamics evolve through a symplectic map. The lift is explicit and algebraic, requiring neither recurrent latent states, transformer decoders, implicit optimization, nor inference-time numerical integration. We instantiate the framework with SympNet predictors and introduce GRB-SympNet, a B-spline variant combining approximation with exact symplectic structure. Experiments on a controlled dissipative double pendulum, a real-world quadrotor, and a contact-constrained real-world quadruped demonstrate the lowest out-of-distribution autoregressive rollout error across systems, improving by up to 69.5% while using fewer parameters and up to 85x fewer floating-point operations per step than sequence-model baselines. The lifted dynamics preserve the symplectic form to numerical precision, extending symplectic learning beyond conservative mechanics toward real-world robotics.