Matt Barnes

Markus Wulfmeier, Michael Bloesch, Nino Vieillard et al. · deepmind

The majority of language model training builds on imitation learning. It covers pretraining, supervised fine-tuning, and affects the starting conditions for reinforcement learning from human feedback (RLHF). The simplicity and scalability of maximum likelihood estimation (MLE) for next token prediction led to its role as predominant paradigm. However, the broader field of imitation learning can more effectively utilize the sequential structure underlying autoregressive generation. We focus on investigating the inverse reinforcement learning (IRL) perspective to imitation, extracting rewards and directly optimizing sequences instead of individual token likelihoods and evaluate its benefits for fine-tuning large language models. We provide a new angle, reformulating inverse soft-Q-learning as a temporal difference regularized extension of MLE. This creates a principled connection between MLE and IRL and allows trading off added complexity with increased performance and diversity of generations in the supervised fine-tuning (SFT) setting. We find clear advantages for IRL-based imitation, in particular for retaining diversity while maximizing task performance, rendering IRL a strong alternative on fixed SFT datasets even without online data generation. Our analysis of IRL-extracted reward functions further indicates benefits for more robust reward functions via tighter integration of supervised and preference-based LLM post-training.

Matt Barnes, Matthew Abueg, Oliver F. Lange et al.

Inverse reinforcement learning (IRL) offers a powerful and general framework for learning humans' latent preferences in route recommendation, yet no approach has successfully addressed planetary-scale problems with hundreds of millions of states and demonstration trajectories. In this paper, we introduce scaling techniques based on graph compression, spatial parallelization, and improved initialization conditions inspired by a connection to eigenvector algorithms. We revisit classic IRL methods in the routing context, and make the key observation that there exists a trade-off between the use of cheap, deterministic planners and expensive yet robust stochastic policies. This insight is leveraged in Receding Horizon Inverse Planning (RHIP), a new generalization of classic IRL algorithms that provides fine-grained control over performance trade-offs via its planning horizon. Our contributions culminate in a policy that achieves a 16-24% improvement in route quality at a global scale, and to the best of our knowledge, represents the largest published study of IRL algorithms in a real-world setting to date. We conclude by conducting an ablation study of key components, presenting negative results from alternative eigenvalue solvers, and identifying opportunities to further improve scalability via IRL-specific batching strategies.

Luyang Liu, David Racz, Kara Vaillancourt et al.

Road crashes are the sixth leading cause of lost disability-adjusted life-years (DALYs) worldwide. One major challenge in traffic safety research is the sparsity of crashes, which makes it difficult to achieve a fine-grain understanding of crash causations and predict future crash risk in a timely manner. Hard-braking events have been widely used as a safety surrogate due to their relatively high prevalence and ease of detection with embedded vehicle sensors. As an alternative to using sensors fixed in vehicles, this paper presents a scalable approach for detecting hard-braking events using the kinematics data collected from smartphone sensors. We train a Transformer-based machine learning model for hard-braking event detection using concurrent sensor readings from smartphones and vehicle sensors from drivers who connect their phone to the vehicle while navigating in Google Maps. The detection model shows superior performance with a $0.83$ Area under the Precision-Recall Curve (PR-AUC), which is $3.8\times$better than a GPS speed-based heuristic model, and $166.6\times$better than an accelerometer-based heuristic model. The detected hard-braking events are strongly correlated with crashes from publicly available datasets, supporting their use as a safety surrogate. In addition, we conduct model fairness and selection bias evaluation to ensure that the safety benefits are equally shared. The developed methodology can benefit many safety applications such as identifying safety hot spots at road network level, evaluating the safety of new user interfaces, as well as using routing to improve traffic safety.

Ethan K. Gordon, Xiang Meng, Matt Barnes et al.

A successful robot-assisted feeding system requires bite acquisition of a wide variety of food items. It must adapt to changing user food preferences under uncertain visual and physical environments. Different food items in different environmental conditions require different manipulation strategies for successful bite acquisition. Therefore, a key challenge is how to handle previously unseen food items with very different success rate distributions over strategy. Combining low-level controllers and planners into discrete action trajectories, we show that the problem can be represented using a linear contextual bandit setting. We construct a simulated environment using a doubly robust loss estimate from previously seen food items, which we use to tune the parameters of off-the-shelf contextual bandit algorithms. Finally, we demonstrate empirically on a robot-assisted feeding system that, even starting with a model trained on thousands of skewering attempts on dissimilar previously seen food items, $ε$-greedy and LinUCB algorithms can quickly converge to the most successful manipulation strategy.