🤖 AI

SPARC: Single-Pass Scaling for Motion Forecasting with Conformal Bayesian Last Layers

The paper introduces SPARC, a single-pass Bayesian-conformal framework for motion forecasting that efficiently generates calibrated, structured uncertainty estimates by combining a deterministic MLP backbone with an analytic Bayesian last layer and split conformal calibration, achieving state-of-the-art performance in both accuracy and uncertainty quantification without requiring Monte Carlo sampling.

Original authors: Sakif Hossain, Julian Teusch, Jörg P. Müller

Published 2026-08-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Sakif Hossain, Julian Teusch, Jörg P. Müller

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

Predicting how a person will move next is a fundamental challenge for machines that interact with the physical world. Whether a robot is handing a tool to a worker, an autonomous vehicle is navigating a crowded street, or a digital character is animating a scene, the system must not only guess the future position of a limb but also understand how sure it is about that guess. In the world of human motion forecasting, researchers have long focused on making the average prediction as accurate as possible, measuring success by how close the predicted joint lands to the actual joint. However, a single point of accuracy is often insufficient for safety. A system needs to know when it is guessing blindly, such as when a person moves in a way the machine has never seen before, or when multiple different futures are equally possible. This requires a way to measure uncertainty that is both structured—respecting the fact that a shoulder moving implies an elbow moving—and reliable, ensuring that the system's confidence matches reality.

For years, the standard way to estimate this uncertainty involved running the same prediction model many times with slight random variations, a process that is slow and computationally expensive. Another approach simply adjusted the final output to guarantee a certain level of coverage, but this often failed to distinguish between the natural randomness of human movement and the model's own lack of knowledge. A new approach, called SPARC, offers a different path. Developed by researchers at the Clausthal University of Technology, this method allows a computer to generate a single, fast prediction that includes a built-in, mathematically rigorous measure of uncertainty. It achieves this without the need for repeated, time-consuming simulations, making it suitable for real-time applications where speed is critical.

The core of this innovation lies in how the researchers modified the final step of a standard motion prediction network. Imagine a machine that has learned to predict a person's future pose by analyzing a short video clip of their past movement. In the new system, the machine first produces a standard, deterministic prediction of where the joints will go. Then, instead of stopping there, it applies a specialized mathematical layer that looks at the specific features of the input to determine how much the model should "trust" its own prediction. If the input looks very similar to the data the model was trained on, the system remains confident. If the input is unusual or falls into a gap in the training data, the system automatically widens its prediction range to account for this lack of knowledge. This adjustment happens instantly, in a single pass through the network, and it inflates the uncertainty in a way that respects the physical connections of the human body, ensuring that the uncertainty of a hand is linked to the uncertainty of the arm.

To ensure these uncertainty estimates are not just theoretical but actually reliable in practice, the researchers added a second step called calibration. This process uses a separate set of data to fine-tune the width of the prediction ranges. The goal is to produce a "tube" of possible future paths around the main prediction. The researchers tested this system across nine different datasets containing thousands of motion capture recordings, ranging from indoor laboratory settings to complex human-robot interactions. They found that the new method produced prediction tubes that were both accurate and efficient. In statistical terms, the system achieved a 95% coverage rate, meaning that in 95 out of 100 cases, the actual future movement of the person fell within the predicted range. Crucially, it did this while maintaining a high level of accuracy for the main prediction itself, outperforming other methods that either sacrificed speed for accuracy or accuracy for reliability.

One of the most practical findings from the study is that the system's internal measure of uncertainty can act as a warning signal. The researchers discovered that they could rank different moments in a video based on how uncertain the model was. When the system flagged a high level of uncertainty, the actual errors in the prediction were significantly larger than average. This suggests that in a real-world application, a robot could use this signal to decide when to slow down, ask for human help, or switch to a safer, more conservative mode of operation. The method successfully separates the natural unpredictability of human movement from the model's own ignorance, providing a clear, single-pass signal that tells a machine when it is on solid ground and when it is guessing.

The study confirms that it is possible to have a motion forecasting system that is fast, accurate, and honest about its limitations without needing complex, slow simulations. By combining a standard prediction engine with a specialized uncertainty layer and a final calibration step, the researchers created a tool that can be deployed in environments where safety and speed are equally important. The work demonstrates that machines can be taught to recognize the boundaries of their own knowledge, a capability that is essential for the next generation of intelligent systems that must interact safely and effectively with people.

Technical Summary: SPARC – Single-Pass Scaling for Motion Forecasting with Conformal Bayesian Last Layers

Problem Statement
Human motion forecasting aims to predict future 3D pose trajectories from short observed prefixes, a critical component for human-robot interaction, animation, and anticipatory perception. While recent deterministic forecasters achieve high accuracy and speed (low Mean Per-Joint Position Error, MPJPE), they lack reliable uncertainty estimates necessary for downstream decision-making. Existing uncertainty quantification methods face significant trade-offs:

Bayesian/Ensemble methods (e.g., MC dropout, deep ensembles) often require repeated stochastic forward passes, making them computationally expensive and potentially unstable under distribution shifts.
Conformal prediction provides finite-sample marginal validity but typically does not separate epistemic uncertainty (model ignorance) from aleatoric uncertainty (intrinsic noise) and often fails to preserve the structured spatio-temporal covariance required for coherent trajectory tubes.
Neural-linear approaches offer analytic uncertainty but frequently rely on factorized (diagonal) observation models, which cannot capture the complex skeletal and temporal correlations inherent in human motion.

Methodology: SPARC
The authors propose SPARC (Single-Pass Adaptive Risk Calibration), a hybrid framework that integrates a deterministic motion forecaster with a Bayesian last layer and post-hoc conformal calibration. The method operates in a single forward pass without Monte Carlo sampling.

Deterministic Backbone: SPARC utilizes a strong deterministic MLP backbone (specifically siMLPe with DCT-bridge) to predict the future mean trajectory $\hat{\mu}_t(x)$ .
Structured Aleatoric Covariance (MN-GraphJ): Instead of assuming independent noise, the model employs a structured Gaussian head to model aleatoric uncertainty. This uses a Matrix-Normal parameterization where the residual covariance is factorized into a temporal component ( $\Sigma_T$ ) and a skeletal component ( $\Sigma_C$ ). The skeletal component is modeled via a Graphical Markov Random Field (GMRF) using a joint-graph precision matrix to encode kinematic dependencies.
Conjugate Bayesian Last Layer: A conjugate Bayesian model is placed on the final linear layer of the backbone. This maps time-domain feature leverage vectors $\phi_t(x)$ $ϕ_{t} (x)$ to an analytic, input-dependent epistemic scale $\kappa_t(x)$ $κ_{t} (x)$ .
- Key Theoretical Result (Prop. 1): Under a matrix-normal prior, integrating out the last-layer weights results in a predictive distribution where the structured covariance $\Sigma_{str,t}$ is preserved exactly but multiplicatively inflated by a scalar factor: $\kappa_t(x) = 1 + \phi_t(x)^\top \Lambda_{n,t}^{-1} \phi_t(x)$ . This allows the injection of epistemic uncertainty without destroying the learned spatio-temporal correlation structure.
- Stabilization: To ensure robustness, the scale is optionally shrunk toward a global sequence-level scale ( $\bar{\kappa}$ ) and can be tuned via an "epistemic temperature" parameter $\gamma$ .
Split Conformal Calibration: To guarantee target coverage under exchangeability, the hybrid Gaussian predictor is wrapped in a split conformal layer. Standardized residuals are computed on a held-out calibration set to derive quantile factors $q_{t,j}$ . These factors convert the probabilistic predictions into 95% marginal prediction tubes (intervals) with finite-sample validity.
Deployment: The final output is a calibrated prediction tube defined by the mean, the hybrid covariance (inflated by $\kappa$ ), and the conformal quantiles. The system also supports $\kappa$ -conditioned (Mondrian) tubes and trajectory-level conformal ellipsoids.

Key Contributions

Structured Bayesian Last-Layer Scaling: The authors derive an exact factorization where epistemic uncertainty inflates a structured covariance matrix via a scalar multiplier, preserving arbitrary positive-definite trajectory structures (dense, Kronecker-factored, or sparse).
Single-Pass Efficiency: The method combines Bayesian epistemic scaling and conformal calibration into a single inference pass, avoiding the computational cost of ensembles or MC sampling.
Cross-Dataset Probabilistic Gains: Across nine dataset/protocol blocks (including Human3.6M, AMASS, and HRI datasets), SPARC achieves the top rank in Negative Log-Likelihood (NLL) and the combined MPJPE+NLL criterion, while maintaining competitive point accuracy and efficient calibrated tube widths.
Deployment-Oriented Risk Monitoring: The epistemic scale $\kappa$ serves as a lightweight risk monitor. Ranking prediction windows by $\kappa$ effectively separates high-error cases; filtering the highest- $\kappa$ decile reduces MPJPE by approximately 9% on Human3.6M.

Results

Performance: SPARC ranks first in average NLL and the combined MPJPE+NLL metric across diverse datasets. It outperforms deterministic baselines, sampling-based ensembles, and calibration-only methods (like Bridge-CQR).
Calibration: The method produces 95% marginal prediction tubes that are both valid (meeting coverage targets) and efficient (narrower than indiscriminately widened intervals).
Ablation: Experiments confirm that the combination of hybrid structured covariance (MN-GraphJ) and time-coupled $\kappa_t$ shrinkage yields the best trade-off between point accuracy, density quality, and tube efficiency.
Robustness: The system remains stable across different random seeds for data splits, model fitting, and hyperparameter optimization.

Significance and Claims
The paper claims that SPARC addresses the critical gap between high-accuracy point forecasting and reliable uncertainty quantification in motion prediction. By integrating analytic Bayesian scaling with conformal guarantees, SPARC provides:

Interpretable Epistemic Scaling: The $\kappa$ signal explicitly quantifies model uncertainty based on feature leverage, allowing for risk-aware decision-making (e.g., triggering fallbacks in safety-critical HRI).
Structured Uncertainty: Unlike methods that assume diagonal noise, SPARC preserves the complex skeletal and temporal correlations essential for realistic trajectory modeling.
Practical Deployability: The single-pass nature and post-hoc calibration make it suitable for real-time applications where computational resources are constrained and reliability is paramount.

The authors note limitations, including the dependence on the quality of the learned feature representation and the fact that conformal guarantees are marginal rather than conditional, though the method offers a practical balance for deployment.

Technical Summary: SPARC – Single-Pass Scaling for Motion Forecasting with Conformal Bayesian Last Layers

More like this