Pre-trained models trained with internet-scale data have achieved significant improvements in perception, interaction, and reasoning. Using them as the basis of embodied grasping methods has greatly promoted the development of robotics applications. In this paper, we provide a comprehensive review of the latest developments in this field. First, we summarize the embodied foundations, including cutting-edge embodied robots, simulation platforms, publicly available datasets, and data acquisition methods, to fully understand the research focus. Then, the embodied algorithms are introduced, starting from pre-trained models, with three main research goals: (1) embodied perception, using data captured by visual sensors to perform point cloud extraction or 3D reconstruction, combined with pre-trained models, to understand the target object and external environment and directly predict the execution of actions; (2) embodied strategy: In imitation learning, the pre-trained model is used to enhance data or as a feature extractor to enhance the generalization ability of the model. In reinforcement learning, the pre-trained model is used to obtain the optimal reward function, which improves the learning efficiency and ability of reinforcement learning; (3) embodied agent: The pre-trained model adopts hierarchical or holistic execution to achieve end-to-end robot control. Finally, the challenges of the current research are summarized, and a perspective on feasible technical routes is provided.

Dual-arm robot manipulation tasks are highly complex, and different tasks can be categorized as dual-arm synchronous coordination or asynchronous sequential execution. Existing action-generative policies adopt a fully parameter-sharing single-network architecture in multi-task learning, overlooking the differences between these two task types in terms of temporal dependencies and action distributions, which leads to cross-mode interference. To address this limitation, we propose LGDM (Language-Guided Dual-Mode Policy), a language-guided dual-mode policy framework. Under shared language embeddings, it parallelly constructs dual-conditional branches for language routing and visual perception. The execution layer selects independent synchronous or asynchronous policy branches according to routing signals and outputs actions by combining them with perceptual features, thereby explicitly decoupling the two motion modes. Specifically, the model organizes three functional modules around language embeddings as a hub: (1) a language mode routing layer that predicts motion modes from temporal information in semantics and provides mode-selection signals to the execution layer; (2) a vision perception layer that injects semantic information into the visual backbone via FiLM, enabling task-aware dynamic feature modulation; and (3) a dual-mode execution layer that builds independent synchronous and asynchronous policy branches, selects the corresponding branch according to routing, and generates actions with the modulated features. On the RoboTwin 2 dual-arm manipulation benchmark comprising six synchronous and asynchronous tasks, LGDM outperforms existing baselines, achieving an absolute improvement of 14.5 percentage points over RDT, and maintains robust execution performance under unseen temporal-coordination instruction variants at inference time.

Robotic manipulation policies have made significant progress in recent years, yet their target-cognitive generalization capability remains insufficient when facing unseen targets and scenarios with similar distractors. Existing methods mostly rely on implicit alignment between language descriptions and global visual features. When target appearance or geometric shape changes, or when similar distractors are present, they struggle to stably establish the correspondence between the language-specified target and action generation, thereby affecting manipulation success rates. To address this problem, this paper proposes TCG-BP (Target-Cognitive Generalization Bimanual Policy), a target-prior-driven bimanual manipulation policy. The method converts language target descriptions into temporally consistent pixel-level target masks, and enhances visual representations through image–mask collaborative encoding and fusion. In the action generation stage, the global scene representation and target-focused representation are extracted from the enhanced visual representations and injected into the policy network in a differentiated manner, enabling continuous action prediction to be constrained by scene context while being guided by target priors. On the RoboTwin 2.0 benchmark, TCG-BP improves the average success rate over π0 by 10.2, 12.2, and 13.8 percentage points under the Seen, Unseen Object, and Unseen Distractor settings, respectively. Experimental results verify the effectiveness of the proposed method.