🤖 machine learning

Conformal Orbit-Valid Trust Horizons for Equivariant World Models

This paper proposes a conformal certification framework for latent world models with known symmetries, demonstrating that exact equivariance enables the transport of calibrated, non-vacuous trust-horizon curves across group orbits while achieving zero anti-conservative violations in empirical audits.

Original authors: Hongbo Wang

Published 2026-06-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Hongbo Wang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Idea: How Long Can You Trust a Robot's Crystal Ball?

Imagine you have a robot that tries to predict the future. It looks at the world, makes a guess about what will happen next, and then uses that guess to plan its next move. This is called a "world model."

The problem is that these robots aren't perfect. If they predict the future too far ahead, their guesses get messy and wrong. If they act on a bad guess, they might crash or drop something.

The Core Question: How far into the future can this robot safely look before its predictions become unreliable?

This paper introduces a "Trust Horizon Certificate." Think of it as a safety expiration date for the robot's predictions. It doesn't just guess; it mathematically proves, "You can trust this prediction for exactly X steps, but no further."

The Secret Weapon: Symmetry (The "Orbit" Concept)

The robots in this study have a special superpower: Symmetry.

Imagine a robot playing with a ball on a perfectly round, frictionless table.

If the robot pushes the ball to the North, it learns how the ball rolls.
Because the table is round and the physics are the same everywhere, the robot should know exactly how the ball would roll if pushed to the East, South, or West. It's the same game, just rotated.

In math terms, this is called Equivariance. The robot understands that the rules of the game don't change just because you turn the table.

The Paper's Big Discovery:
Usually, to be safe, a robot needs to test its predictions in every direction (North, East, South, West) to make sure it works everywhere. That takes a lot of time and data.

But because this robot understands symmetry, the authors found a shortcut:

If you prove the robot is safe in one direction (a "wedge"), you automatically know it is safe in all directions.

It's like testing a tire on a car. If you prove the tire holds up when the car drives North, and you know the car is perfectly symmetrical, you don't need to test it driving East, South, or West. The "Trust Certificate" you made for the North trip is valid for the whole trip around the world.

How They Built the "Trust Certificate"

The authors used a statistical trick called Conformal Prediction. Here is the analogy:

The Raw Guess: The robot makes a prediction and says, "I think I'm 90% accurate."
The Safety Margin: The authors say, "We don't trust that 90% yet. Let's add a safety buffer." They multiply the error by a factor (like adding a 10% safety margin to a bridge's weight limit).
The Audit: They test the robot 50 times. In every single test, the robot stayed within the safety margin.
The Result: They created a certificate that says, "We are 95% sure this robot won't fail within this time limit."

The Magic: Because of the symmetry (the "Orbit" concept), they only had to do this testing on a small slice of the world. Once that slice was certified, the math proved the entire circle was certified for free.

The Two Worlds: Where It Works and Where It Doesn't

The paper tested this in two different environments, and the results were surprisingly nuanced:

1. The Round Table (2D Symmetric World)

The Setup: A flat, perfectly round surface where everything looks the same from every angle.
The Result: Even robots without the symmetry superpower (the "plain" robots) worked fine here. Because the world was so simple and round, the plain robots accidentally figured out the rules just by seeing enough examples.
The Lesson: In a perfectly symmetrical world, having a symmetry superpower doesn't save you much time because the world is already easy to learn.

2. The 3D Tabletop (The Real World)

The Setup: A 3D table where gravity pulls down. You can spin the table left and right (yaw), but you can't flip it upside down (gravity breaks the full symmetry).
The Result: Here, the symmetry superpower mattered.
- The Symmetric Robot got a "One-Step Certificate." It only needed to test one direction to be safe everywhere else.
- The Plain Robot failed. It had to test many directions, or it kept making mistakes. It either had to be overly cautious (wasting time) or it broke the safety rules.
The Lesson: In complex, real-world scenarios where symmetry exists but isn't perfect, the superpower saves you a massive amount of testing time and risk.

The Catch: The "Planning" Problem

The authors tried to use this certificate to help the robot plan its moves (like deciding, "I will aim for a goal 3 steps away").

The Hope: "If the certificate says I'm safe for 3 steps, I'll plan 3 steps ahead."
The Reality: The robot's internal "brain" (the predictor) was a bit shaky. Even though the certificate said "Safe," the robot's actual performance varied wildly from one attempt to another.
The Conclusion: The certificate is a great safety guardrail, but it didn't automatically make the robot a better planner in this specific experiment. The robot still struggled to predict exactly how to move its arm to hit a target, even if it knew when to stop.

Summary in One Sentence

This paper proves that if a robot understands the symmetrical rules of its world, you can test its safety in just one direction and mathematically guarantee it is safe in all directions, saving huge amounts of time and data, though this doesn't automatically solve every planning problem the robot faces.

Technical Summary: Conformal Orbit-Valid Trust Horizons for Equivariant World Models

1. Problem Statement

Learned world models are inherently limited by their rollout error, which degrades over time. While model-based planning and self-supervision rely on rolling these models forward, there is no principled method to determine the "trust horizon"—the specific number of steps over which a model's predictions remain reliable. Existing approaches often rely on fixed hyperparameters (e.g., a fixed horizon $H=25$ ) that are specific to a single distribution and lack formal guarantees.

Furthermore, for world models with known group symmetries (equivariant models), a critical question remains: Can a trust certificate derived from a local region be valid across the entire group orbit without requiring separate calibration for every transformed state? Current methods either lack formal safety guarantees or fail to leverage the geometric structure of equivariance to reduce calibration costs.

2. Methodology

2.1 The Equivariant World Model Setup

The authors define a latent world model consisting of:

Encoder ( $E$ ): Maps observations $x$ to latents $z$ .
Predictor ( $f$ ): Advances latents conditioned on actions $a$ ( $\hat{z}_{t+1} = f(z_t, a_t)$ ).
Symmetry: The model satisfies equivariance conditions where $E(g \cdot x) = \rho(g)E(x)$ and $f(\rho(g)z, \sigma(g)a) = \rho(g)f(z,a)$ for a group $g \in \mathcal{G}$ .
Implementation: Encoders are realized as CN-steerable networks (2D) or Vector-Neuron/SE(3) networks (3D). Equivariance is verified via unit tests checking residuals against machine precision or learned tolerances.

2.2 Raw Horizon and Conformal Calibration

The core of the methodology involves constructing a trust horizon certificate:

Raw Curve: A one-step latent residual $\hat{\delta}$ and a finite-time expansion estimate $\hat{\lambda}_1$ (derived from the predictor Jacobian via Benettin/QR schemes) form a raw error propagation curve:
$\widehat{Err}(H) = \hat{\delta} \sum_{t=0}^{H-1} e^{\hat{\lambda}_1 t}$
The raw horizon $H_{raw}$ is the maximum steps where this error stays below a tolerance $\epsilon$ .
Split-Conformal Calibration: To ensure statistical validity, the raw curve is scaled by a multiplicative factor $\gamma_\alpha$ $γ_{α}$ . The certified horizon $H_{conf}$ $H_{co n f}$ is defined such that $\widehat{Err}_{conf}(H) = \gamma_\alpha \widehat{Err}(H) \le \epsilon$ $E r r_{co n f} (H) = γ_{α} E r r (H) \leq ϵ$ .
- The factor $\gamma_\alpha$ is determined via split-conformal prediction on a calibration set, ensuring that for a fresh exchangeable audit, $P(H_{conf} \le H_{meas}) \ge 1-\alpha$ .
- Crucially, in the authors' reproducible audit set, the raw curve was already conservative, resulting in $\gamma_\alpha = 1.0$ .

2.3 Orbit-Valid Transport

The paper introduces a structural result combining the statistical calibration with the geometric property of equivariance:

Theorem A (Orbit Transport): If the model is exactly equivariant, rollout errors, expansion rates, and the resulting horizon curves are constant along the group orbit.
Corollary C (Orbit-Valid Certificate): If a certificate is calibrated on a fundamental region (a "wedge") and exact equivariance holds, the certificate is valid for the entire orbit at the same confidence level $1-\alpha$ .
Implication: Equivariance transforms the orbit-wide coverage event from a union over all group transformations (which would require union bounds and massive data) into a single calibrated event. This eliminates the need for orbit-wise calibration data.

2.4 Handling Approximate Equivariance

Since trained networks are only approximately equivariant, the authors account for transport residuals ( $R_t$ ) using Lipschitz bounds on the encoder and predictor defects. Empirically, these residuals were found to be small (median 1.1%), allowing the certificate to remain valid with minimal inflation.

3. Key Contributions

Conformal Trust-Horizon Certificate: A method to convert a raw latent error-propagation curve into a one-sided, conservative certificate using split-conformal calibration. The method achieved 0/50 anti-conservative violations on stable audits, with an exact-binomial 95% upper bound of 5.8% on the violation rate.
Orbit-Validity Theorem: A proof that exact equivariance allows a certificate calibrated on a single sector to be transported over the entire group orbit without additional calibration cost (Corollary C).
Sharpness and Residual Accounting: The certificate is non-vacuous, with a median certified-to-measured horizon ratio ( $H_{conf}/H_{meas}$ ) of 0.67. Implementation-level transport residuals were small (median 1.1%, max 4.1%).
Substrate-Dependent Calibration Cost: The paper identifies that the benefit of equivariance in reducing calibration cost is conditional on the data substrate.
- 2D Symmetric Substrate: Non-equivariant baselines were already approximately orbit-robust due to the symmetric environment, so equivariance provided no separation in calibration cost.
- 3D Yaw Substrate: Non-equivariant baselines failed to achieve orbit safety without significant slack, violations, or seed instability. Equivariant models achieved a safe, non-vacuous certificate from a single calibration sector.

4. Results and Empirical Findings

Safety: Across 50 stable audit cells, the conformal certificate with $\gamma_\alpha=1.0$ showed zero anti-conservative violations.
Sharpness: In 172 cell-level checks, 98% of checks retained at least 25% of the measured horizon, and 69% retained at least 50%. The certificate acts as a sharp feasibility gate for short horizons ( $H \in [1,8]$ ).
Calibration Cost:
- In 2D (PushT), equivariant, plain, and augmented models were all orbit-valid from one sector.
- In 3D (ManiSkill yaw), the equivariant model obtained a one-sector safe certificate, while healthy non-equivariant baselines incurred costs in the form of violations, slack multipliers ( $\gamma_\alpha \approx 1.9$ ), or sharpness collapse.
Planning Application (Negative Result): The authors attempted to use the certified horizon to guide subgoal spacing in a 3D CEM-MPC planner. This did not succeed. The planning substrate exhibited high variance across seeds, and no detectable association was found between the certified trust horizon and the optimal planning horizon. The authors attribute this to a "state-sufficiency bottleneck," where the latent representation lacked velocity information necessary for dynamics prediction in a gravity-broken environment.

5. Significance and Claims

The paper claims that equivariance does not replace calibration; it gives calibration a geometry.

Structural vs. Empirical: The orbit-transport theorem is a structural guarantee that holds whenever exact equivariance is present. However, the empirical benefit (reduced calibration cost) only manifests when the data distribution does not already induce approximate orbit robustness in non-equivariant baselines.
Nature of the Guarantee: The certificate is a conservative, distributional audit, not a global reachability guarantee. It certifies that the model is trustworthy relative to the measured audit distribution, not that it is safe in all possible worst-case scenarios.
Modesty on Planning: The authors explicitly state that the "certified subgoal spacing" application was not confirmed. The negative result in the 3D planner is reported as a boundary finding, highlighting that the structure-variance economics of equivariant models (high variance in planning performance) can negate theoretical benefits in downstream control layers.

The work distinguishes itself from prior art by calibrating a horizon function rather than a scalar or interval, and by transporting that function over a group orbit, thereby converting a local audit into an orbit-valid certificate. It avoids claiming that equivariance universally improves prediction accuracy or planning performance, focusing instead on the geometric efficiency of the certification process itself.