Chendi Qu

Yi Wang, Xinchen Li, Pengwei Xie et al.

Generalist robot policies increasingly benefit from large-scale pretraining, but offline data alone is insufficient for robust real-world deployment. Deployed robots encounter distribution shifts, long-tail failures, task variations, and human correction opportunities that fixed demonstration datasets cannot fully capture. We present Learning While Deploying (LWD), a fleet-scale offline-to-online reinforcement learning framework for continual post-training of generalist Vision-Language-Action (VLA) policies. Starting from a pretrained VLA policy, LWD closes the loop between deployment, shared physical experience, policy improvement, and redeployment by using autonomous rollouts and human interventions collected across a robot fleet. To stabilize learning from heterogeneous, sparse-reward fleet data, LWD combines Distributional Implicit Value Learning (DIVL) for robust value estimation with Q-learning via Adjoint Matching (QAM) for policy extraction in flow-based VLA action generators. We validate LWD on a fleet of 16 dual-arm robots across eight real-world manipulation tasks, including semantic grocery restocking and 3--5 minute long-horizon tasks. A single generalist policy improves as fleet experience accumulates, reaching an average success rate of 95%, with the largest gains on long-horizon tasks.

Chenlin Ming, Chendi Qu, Mengzhang Cai et al.

Large Language Models (LLMs) have achieved impressive performance through Supervised Fine-tuning (SFT) on diverse instructional datasets. When training on multiple capabilities simultaneously, the mixture training dataset, governed by volumes of data from different domains, is a critical factor that directly impacts the final model's performance. Unlike many studies that focus on enhancing the quality of training datasets through data selection methods, few works explore the intricate relationship between the compositional quantity of mixture training datasets and the emergent capabilities of LLMs. Given the availability of a high-quality multi-domain training dataset, understanding the impact of data from each domain on the model's overall capabilities is crucial for preparing SFT data and training a well-balanced model that performs effectively across diverse domains. In this work, we introduce IDEAL, an innovative data equilibrium adaptation framework designed to effectively optimize volumes of data from different domains within mixture SFT datasets, thereby enhancing the model's alignment and performance across multiple capabilities. IDEAL employs a gradient-based approach to iteratively refine the training data distribution, dynamically adjusting the volumes of domain-specific data based on their impact on downstream task performance. By leveraging this adaptive mechanism, IDEAL ensures a balanced dataset composition, enabling the model to achieve robust generalization and consistent proficiency across diverse tasks. Experiments across different capabilities demonstrate that IDEAL outperforms conventional uniform data allocation strategies, achieving a comprehensive improvement of approximately 7% in multi-task evaluation scores.

Chendi Qu, Jianping He, Xiaoming Duan et al.

Inverse reinforcement learning (IRL) usually assumes the reward function model is pre-specified as a weighted sum of features and estimates the weighting parameters only. However, how to select features and determine a proper reward model is nontrivial and experience-dependent. A simplistic model is less likely to contain the ideal reward function, while a model with high complexity leads to substantial computation cost and potential overfitting. This paper addresses this trade-off in the model selection for IRL problems by introducing the structural risk minimization (SRM) framework from statistical learning. SRM selects an optimal reward function class from a hypothesis set minimizing both estimation error and model complexity. To formulate an SRM scheme for IRL, we estimate the policy gradient from given demonstration as the empirical risk, and establish the upper bound of Rademacher complexity as the model penalty of hypothesis function classes. The SRM learning guarantee is further presented. In particular, we provide the explicit form for the linear weighted sum setting. Simulations demonstrate the performance and efficiency of our algorithm.