# Scaling Laws of Motion Forecasting and Planning -- A Technical Report

**Authors:** Mustafa Baniodeh, Kratarth Goel, Scott Ettinger, Carlos Fuertes, Ari Seff, Tim Shen, Cole Gulino, Chenjie Yang, Ghassen Jerfel, Dokook Choe, Rui Wang, Vinutha Kallem, Sergio Casas, Rami Al-Rfou, Benjamin Sapp, Dragomir Anguelov  
**arXiv:** [2506.08228](https://arxiv.org/abs/2506.08228)  
**Date:** 2025-06-09  
**Summary page:** https://www.summarizepaper.com/en/paper/2506.08228v1/

## Key points

- Researchers studied scaling laws of encoder-decoder autoregressive transformer models in joint motion forecasting and planning for autonomous driving
- Model performance improves with increased compute budget following a power-law relationship
- Significant correlation between model training loss and evaluation metrics
- Closed-loop metrics show enhancement with increasing scaling, impacting model development and optimization
- Optimal scaling parameters for transformer parameters and training data size to achieve a compute-optimal model were determined
- Increasing training compute budget requires model size to increase 1.5 times faster than dataset size for optimal scaling
- Inference-time compute scaling involves sampling and clustering output from smaller models until larger models become more efficient at a crossover point
- Optimizing both training and inference-time scaling properties enhances performance across diverse driving scenarios
- Leveraging general logged driving data from other agents can improve ego-agents' performance, addressing data scarcity challenges in large capacity model training for robotics applications.

## Summary

In their technical report titled "Scaling Laws of Motion Forecasting and Planning," Mustafa Baniodeh, Kratarth Goel, Scott Ettinger, Carlos Fuertes, Ari Seff, Tim Shen, Cole Gulino, Chenjie Yang, Ghassen Jerfel, Dokook Choe, Rui Wang,
Vinutha Kallem,Sergio Casas,Rami Al-Rfou,Benjamin Sapp,and Dragomir Anguelov delve into the empirical scaling laws of encoder-decoder autoregressive transformer models in the context of joint motion forecasting and planning within autonomous driving. Utilizing a vast dataset comprising 500 thousand hours of driving data,the researchers reveal that akin to language modeling tasks,model performance exhibits improvement following a power-law relationship with the total compute budget allocated. They also identify a significant correlation between model training loss and evaluation metrics. Notably,
as scaling increases,closed-loop metrics show enhancement as well - a finding with crucial implications for model development and optimization using open-loop metrics. The study further delves into determining the optimal scaling parameters for transformer parameters
and training data size to achieve a compute-optimal model during training.The team discovers that with an expanding training compute budget,optimal scaling necessitates increasing the model size at a rate 1.5 times faster than that of the dataset size.Additionally explored is inference-time compute scaling where sampling and clustering output from smaller models enable them to compete effectively with larger models until reaching a crossover point where larger models become more efficient in terms of inference-compute trade-offs. Overall experimental results underscore the importance of optimizing both training and inference-time scaling properties in enhancing the performance of motion forecasting and planning models across diverse driving scenarios.Lastly,the report briefly touches upon leveraging general logged driving data from other agents to bolster the performance of ego-agents - addressing a critical research area aimed at mitigating data scarcity challenges in training large capacity models for robotics applications.