Generalist robot manipulation policies (GMPs) have the potential to generalize across a wide range of tasks, devices, and environments. However, existing policies continue to struggle with out-of-distribution scenarios due to the inherent difficulty of collecting sufficient action data to cover extensively diverse domains. While fine-tuning offers a practical way to quickly adapt a GMPs to novel domains and tasks with limited samples, we observe that the performance of the resulting GMPs differs significantly with respect to the design choices of fine-tuning strategies. In this work, we first conduct an indepth empirical study to investigate the effect of key factors in GMPs fine-tuning strategies, covering the action space, policy head, supervision signal and the choice of tunable parameters, where 2,500 rollouts are evaluated for a single configuration. We systematically discuss and summarize our findings and identify the key design choices, which we believe give a practical guideline for GMPs fine-tuning. We observe that in a lowdata regime, with carefully chosen fine-tuning strategies, a GMPs significantly outperforms the state-of-the-art imitation learning algorithms. The results presented in this work establish a new baseline for future studies on fine-tuned GMPs.

We present Chain-of-Action (CoA), a novel visuomotor policy paradigm built upon Trajectory Autoregressive Modeling. Unlike conventional approaches that predict next step action(s) forward, CoA generates an entire trajectory by explicit backward reasoning with task-specific goals through an action-level Chain-of-Thought (CoT) process. This process is unified within a single autoregressive structure: (1) the first token corresponds to a stable keyframe action that encodes the task-specific goals; and (2) subsequent action tokens are generated autoregressively, conditioned on the initial keyframe and previously predicted actions. This backward action reasoning enforces a global-to-local structure, allowing each local action to be tightly constrained by the final goal. To further realize the action reasoning structure, CoA incorporates four complementary designs: continuous action token representation; dynamic stopping for variable-length trajectory generation; reverse temporal ensemble; and multi-token prediction to balance action chunk modeling with global structure. As a result, CoA gives strong spatial generalization capabilities while preserving the flexibility and simplicity of a visuomotor policy. Empirically, we observe that CoA outperforms representative imitation learning algorithms such as ACT and Diffusion Policy across 60 RLBench tasks and 8 real-world tasks.

Large Vision-Language Models (LVLMs) show significant strides in general-propose multimodal applications such as visual dialogue and embodied navigation. However, existing multimodal evaluation benchmarks cover a limited number of multimodal tasks testing rudimentary capabilities, falling short in tracking LVLM development. In this study, we present MMT-Bench, a comprehensive benchmark designed to assess LVLMs across massive multimodal tasks requiring expert knowledge and deliberate visual recognition, localization, and reasoning. MMT-Bench comprises $31,325$ meticulously curated multi-choice visual questions from various multimodal scenarios such as vehicle driving and embodied navigation, covering $32$ core meta-tasks and $162$ subtasks in multimodal understanding. Due to its extensive task coverage, MMT-Bench enables the evaluation of LVLMs using a task map, facilitating the discovery of in- and out-of-domain tasks. Evaluation results involving $20$ publicly available LVLMs such as the proprietary GeminiProVision model, underscore the significant challenges posed by MMT-Bench. We anticipate that MMT-Bench will inspire the community to develop next-generation multimodal foundation models aimed at achieving general-purpose multimodal intelligence.

Depth information has been proved beneficial in RGB-D salient object detection (SOD). However, depth maps obtained often suffer from low quality and inaccuracy. Most existing RGB-D SOD models have no cross-modal interactions or only have unidirectional interactions from depth to RGB in their encoder stages, which may lead to inaccurate encoder features when facing low quality depth. To address this limitation, we propose to conduct progressive bidirectional interactions as early in the encoder stage, yielding a novel bi-directional transfer-and-selection network named BTS-Net, which adopts a set of bi-directional transfer-and-selection (BTS) modules to purify features during encoding. Based on the resulting robust encoder features, we also design an effective light-weight group decoder to achieve accurate final saliency prediction. Comprehensive experiments on six widely used datasets demonstrate that BTS-Net surpasses 16 latest state-of-the-art approaches in terms of four key metrics.

RGB-D salient object detection (SOD) recently has attracted increasing research interest by benefiting conventional RGB SOD with extra depth information. However, existing RGB-D SOD models often fail to perform well in terms of both efficiency and accuracy, which hinders their potential applications on mobile devices and real-world problems. An underlying challenge is that the model accuracy usually degrades when the model is simplified to have few parameters. To tackle this dilemma and also inspired by the fact that depth quality is a key factor influencing the accuracy, we propose a novel depth quality-inspired feature manipulation (DQFM) process, which is efficient itself and can serve as a gating mechanism for filtering depth features to greatly boost the accuracy. DQFM resorts to the alignment of low-level RGB and depth features, as well as holistic attention of the depth stream to explicitly control and enhance cross-modal fusion. We embed DQFM to obtain an efficient light-weight model called DFM-Net, where we also design a tailored depth backbone and a two-stage decoder for further efficiency consideration. Extensive experimental results demonstrate that our DFM-Net achieves state-of-the-art accuracy when comparing to existing non-efficient models, and meanwhile runs at 140ms on CPU (2.2x faster than the prior fastest efficient model) with only ~8.5Mb model size (14.9% of the prior lightest). Our code will be available at https://github.com/zwbx/DFM-Net.