Causal-PIK: Causality-based Physical Reasoning with a Physics-Informed Kernel

Source

• Raw Markdown: paper_causal-pik-2025.md
• PDF: paper_causal-pik-2025.pdf
• Preprint: https://arxiv.org/abs/2505.22861
• Official project page: https://stanford-iprl-lab.github.io/causal-pik/
• Peer-reviewed proceedings: https://proceedings.mlr.press/v267/morlans25a.html

The paper was submitted to arXiv on 2025-05-28, revised on 2025-05-30, and published at ICML 2025 in PMLR 267. The official project page and author/lab publication pages were found, but no verified public code repository or official X/Twitter announcement was discoverable during the 2026-08-31 ingest.

Core Claim

Causal-PIK uses a short-horizon learned dynamics model to compare candidate actions by their predicted immediate physical effects. It inserts that action similarity directly as the covariance kernel of a Gaussian-process surrogate over task reward, then uses Bayesian optimization to choose informative and promising actions. On Virtual Tools and PHYRE, the paper reports higher AUCCESS than its RBF-kernel ablation and fewer attempts than the listed full-action-space baselines.

The Exact PIK To GP Connection

The Physics-Informed Kernel (PIK) is not a separate matrix that is later converted into a GP by another learned model. It is the GP covariance function. For attempted actions $X=\{x_1,\dots ,x_n\}$ with observed rewards $y=[y_1,\dots ,y_n{]}^{\top }$, the method evaluates PIK on every pair of actions:

$$K_{ij}=k_{\text{PIK}}(x_i,x_j).$$

For a candidate action $x_{\ast }$, it also computes the covariance vector

$$k_{\ast }={\left[\begin{array}{ccc}{k}_{\text{PIK}}(x_1,x_{\ast })& \cdots & k_{\text{PIK}}(x_n,x_{\ast })\end{array}\right]}^{\top }.$$

Under the standard noisy-GP equations, these quantities propagate observed rewards from attempted actions to an untried action:

$$\mu (x_{\ast })=m(x_{\ast })+k_{\ast }^{\top }(K+{\sigma }_n^{2}I{)}^{-1}(y-m_X),$$ $${\sigma }^2(x_{\ast })=k_{\text{PIK}}(x_{\ast },x_{\ast })-k_{\ast }^{\top }(K+{\sigma }_n^{2}I{)}^{-1}{k}_{\ast }.$$

The paper states the GP/kernel relationship but does not print these posterior equations or disclose the observation-noise and UCB exploration settings. Accordingly, the formulas above explain the standard GP operation implied by the method rather than undocumented implementation details.

The important mechanism is reward transfer. If a failed attempted action $x_i$ and an untried candidate $x_{\ast }$ are predicted to produce similar object motion, $k_{\text{PIK}}(x_i,x_{\ast })$ is large, so the observed reward at $x_i$ can strongly influence the GP posterior at $x_{\ast }$. In the one-observation or simple local case, a low reward lowers the candidate posterior mean in effect space, even if the action vectors are not geometrically close. With multiple observations, however, $(K+{\sigma }_n^{2}I{)}^{-1}$ can produce signed weights, so a monotonic pull-down is not guaranteed. An RBF kernel instead transfers information mainly between actions that are close in raw action coordinates.

flowchart LR
  A[Candidate action x] --> D[Learned dynamics model]
  S[Initial scene state] --> D
  D --> E[Predicted immediate object effects]
  E --> P[Pairwise PIK similarities]
  P --> G[GP covariance K and k_*]
  R[Observed rewards of tried actions] --> G
  G --> M[Posterior mean and uncertainty]
  M --> U[UCB over 500 Sobol candidates]
  U --> T[Top 5 candidates]
  T --> N[Probabilistic physics simulation]
  N --> X[Execute best expected candidate]
  X --> R

How PIK Is Computed

For each action, an RPIN dynamics model predicts $n_{\text{pred}}=20$ future steps, usually enough to cover one collision rather than a full rollout. At the first predicted interaction event, the method measures the later state change of each dynamic object $O$:

$${\dot{s}}^{O,a}=\frac{s_{t_{\text{event}}+\Delta t}^O-s_{t_{\text{event}}}^O}{\Delta t}.$$

For actions $a$ and $b$, the per-object similarity multiplies a clipped cosine similarity by a magnitude similarity:

$$c_O(a,b)=\max \left(0,\frac{{\dot{s}}^{O,a}\cdot {\dot{s}}^{O,b}}{\Vert {\dot{s}}^{O,a}\Vert \Vert {\dot{s}}^{O,b}\Vert }\right),$$ $$r_O(a,b)=\frac{1}{1+\left|\Vert {\dot{s}}^{O,a}\Vert -\Vert {\dot{s}}^{O,b}\Vert \right|},s_O(a,b)=c_O(a,b)r_O(a,b).$$

It averages this score over the $D$ dynamic objects and sharpens the result:

$$\bar{s}(a,b)=\frac{1}{D}\sum _{O=1}^D{s}_O(a,b),k_{\text{PIK}}(a,b)=\bar{s}(a,b)e^{\bar{s}(a,b)-1}.$$

Thus PIK groups actions by the predicted direction and magnitude of their immediate effects, not by raw placement distance.

How The Next Action Is Chosen

Each Bayesian-optimization iteration follows this sequence:

1. Refit or update the GP on all attempted actions and observed rewards using PIK as its covariance function.
2. Draw 500 candidate actions with a Sobol sequence.
3. Score each candidate with an Upper Confidence Bound acquisition function, conventionally of the form $\mu (x)+\sqrt{{\beta }_t}\sigma (x)$.
4. Keep the five candidates with the largest UCB values. The mean term exploits actions predicted to score well; the uncertainty term explores actions whose reward is still poorly known.
5. Run a probabilistic intuitive-physics simulation for those five candidates and choose the one with the highest expected simulated outcome.
6. Execute only that action in the benchmark environment, observe its trajectory and reward, append $(x,y)$ to the GP data, and repeat until success.

The acquisition function therefore does not choose the final action alone. It is a 500-to-5 screening stage, followed by simulation-based reranking. The paper does not specify the UCB exploration coefficient, GP noise treatment, or kernel-conditioning implementation, and no public code was found to recover those details.

Before counted attempts begin, the method also seeds the GP with nine noisy simulation rollouts sampled near scene objects. These warm-up points do not count toward the reported attempt budget.

Evidence And Results

• On Virtual Tools, Causal-PIK reports AUCCESS $65.0\pm 25.0$, versus $42.0\pm 33.0$ for its RBF ablation and $58.0\pm 27.0$ for SSUP.
• On PHYRE-1B Cross, it reports $41.6\pm 9.33$, versus $27.70\pm 19.68$ for the RBF ablation. Some other listed methods use drastically reduced action spaces, whereas Causal-PIK searches the full approximately 2.5M-action space.
• With PHYRE limited to ten attempts, the paper reports Causal-PIK at $24.8\pm 9.22$ and humans at $36.6\pm 10.2$; the broad claim that the model exceeds humans therefore depends on the attempt protocol.

The RBF ablation is the clearest evidence for the PIK contribution, but it does not isolate every component of the hybrid pipeline: both variants still use GP Bayesian optimization, warm-up simulations, UCB screening, and simulation-based reranking.

Mathematical And Reproducibility Caveat

The paper’s validity argument for PIK as a GP kernel is incorrect as written. It observes that the similarity is symmetric and that every pairwise value is non-negative, then concludes that the kernel is positive semi-definite (PSD). Entrywise non-negativity does not imply that every finite Gram matrix is PSD.

The issue is not merely formal. For one object with eight equal-magnitude effect vectors equally spaced around the unit circle, evaluating the paper’s own $k_{\text{PIK}}$ formula gives a circulant Gram matrix with a negative eigenvalue of approximately $-0.0551$. Specifically, its first row is $[1,a,0,0,0,0,0,a]$, where $a=2^{-1/2}{e}^{2^{-1/2}-1}\approx 0.5275736$; the alternating Fourier eigenvalue is $1-2a\approx -0.0551473$. Therefore the proposed similarity is not a valid GP covariance kernel in full generality unless additional constraints or a PSD repair are applied.

This matters because exact GP inference normally requires a PSD covariance matrix for stable Cholesky factorization and non-negative posterior variances. The paper and available official pages do not say whether the implementation adds jitter, clips eigenvalues, projects the matrix to the PSD cone, or happens to avoid indefinite cases on its benchmark data. Without public code, the gap cannot be resolved.

Limitations

• The learned effect predictor can produce misleading similarities; the paper identifies this noise as a direct performance limitation.
• Experiments cover low-dimensional single-intervention 2D puzzle tasks, not long-horizon feedback control or real robotic deployment.
• The dynamics model predicts immediate effects, usually through one collision, rather than a complete action-conditioned trajectory.
• The term causal refers primarily to predicted action-attributable effects and a no-action baseline. The work does not establish causal discovery or identification from confounded observational trajectories.
• No verified public code was found, and GP/UCB settings required for exact reproduction are omitted from the paper.

Foundation TSFM Relevance

Agenda slot	Verdict	Evidence	Missing pieces
Control and counterfactuals	partially closes	Uses explicit single-step interventions, observed rewards, and sequential candidate-action selection to reduce physical interactions.	No long-horizon control-input sequence, calibrated counterfactual trajectory distribution, or real-system validation.
Causal structure	adjacent	Compares actions by dynamics-predicted immediate effects rather than raw action distance.	No causal discovery or identification from confounded trajectories; PIK’s PSD claim is false in general.
Latent-state prediction	adjacent	RPIN predicts short post-action object trajectories and supplies decision-relevant effect features.	Prediction is normally limited to one collision and is used for similarity, not a maintained latent state or full rollout.
Benchmarks and evaluation hygiene	warning	GP-UCB screens 500 candidates, then probabilistic simulation reranks five before execution; the RBF ablation tests the effect-space kernel choice.	UCB/GP implementation details and public code are absent; the PIK covariance is indefinite in general, and no high-dimensional or receding-horizon evaluation is reported.

Links Into The Wiki

• World Models
• Causal Time Series
• Robotics Time-Series Modeling
• Time-Series Benchmark Hygiene
• Foundation Time-Series Model Research Agenda

Open Questions

• Can the action-effect similarity be replaced by a provably PSD kernel, for example through an explicit feature map or a PSD projection, without losing the reported sample-efficiency gain?
• How much of the gain comes from effect-space GP generalization versus the top-five probabilistic-simulation reranker?
• Does calibrated GP uncertainty survive dynamics-model errors and distribution shift, or does PIK become confidently wrong when predicted effects are wrong?
• Can this pattern scale from one-shot action placement to action sequences and receding-horizon control inputs?