Scene-LLM is a 3D-visual-language model that enhances the ability of embodied agents to interact with 3D indoor environments by integrating the reasoning capabilities of large language models (LLMs). The model uses a hybrid 3D visual feature representation that combines dense spatial information and supports scene state updates. It employs a projection layer to efficiently project these features into the pre-trained textual embedding space, enabling effective interpretation of 3D visual information. The model integrates both scene-level and egocentric 3D information, which is crucial for interactive planning. Scene-level data supports global planning, while egocentric data is important for localization. The model uses egocentric 3D frame features for feature alignment, an efficient technique that incorporates fine-grained concepts. Experiments with Scene-LLM demonstrate its strong capabilities in dense captioning, question answering, and interactive planning. Scene-LLM advances the field of 3D visual understanding and reasoning, offering new possibilities for sophisticated agent interactions in indoor settings. The model introduces a 3D visual representation that captures fine-grained 3D information and supports state changes by design. This representation can be easily incorporated into LLMs with a lightweight projector. The model also creates a large-scale dataset with 190k 3D-visual-language pairs from an egocentric viewpoint and about 500k pairs of scene-level data. Scene-LLM is evaluated on benchmarks such as ScanQA and SQA3D, achieving state-of-the-art results without additional fine-tuning. It excels in 3D scene reasoning tasks and outperforms other LLM-based models on the Alfred benchmark. Scene-LLM's hybrid feature representation effectively captures comprehensive spatial information and facilitates dynamic state updates. The model's performance is further enhanced by incorporating 3D frame data and using a two-stage training strategy. Scene-LLM's ability to understand egocentric and scene-level information enables it to handle scene changes and perform interactive planning in dynamic environments. The model's performance is validated through experiments on various tasks, demonstrating its effectiveness in 3D visual understanding and reasoning. Scene-LLM's contributions include introducing a 3D-VLM that connects 3D visual information with LLMs and achieving state-of-the-art results on 3D-VQA and interactive planning benchmarks. The model also proposes an effective 3D visual representation that supports state changes and can be easily integrated into LLMs. The model creates a large-scale dataset for 3D and text feature alignment, which includes 190k 3D-visual-language pairs from an egocentric viewpoint and about 500k pairs of scene-level data. Scene-LLM's performance is evaluated on various benchmarks, showing its effectiveness in 3D visual understanding and reasoning. The model's ability to understand egocentric and scene-level information enables it to handle scene changes and perform interactiveScene-LLM is a 3D-visual-language model that enhances the ability of embodied agents to interact with 3D indoor environments by integrating the reasoning capabilities of large language models (LLMs). The model uses a hybrid 3D visual feature representation that combines dense spatial information and supports scene state updates. It employs a projection layer to efficiently project these features into the pre-trained textual embedding space, enabling effective interpretation of 3D visual information. The model integrates both scene-level and egocentric 3D information, which is crucial for interactive planning. Scene-level data supports global planning, while egocentric data is important for localization. The model uses egocentric 3D frame features for feature alignment, an efficient technique that incorporates fine-grained concepts. Experiments with Scene-LLM demonstrate its strong capabilities in dense captioning, question answering, and interactive planning. Scene-LLM advances the field of 3D visual understanding and reasoning, offering new possibilities for sophisticated agent interactions in indoor settings. The model introduces a 3D visual representation that captures fine-grained 3D information and supports state changes by design. This representation can be easily incorporated into LLMs with a lightweight projector. The model also creates a large-scale dataset with 190k 3D-visual-language pairs from an egocentric viewpoint and about 500k pairs of scene-level data. Scene-LLM is evaluated on benchmarks such as ScanQA and SQA3D, achieving state-of-the-art results without additional fine-tuning. It excels in 3D scene reasoning tasks and outperforms other LLM-based models on the Alfred benchmark. Scene-LLM's hybrid feature representation effectively captures comprehensive spatial information and facilitates dynamic state updates. The model's performance is further enhanced by incorporating 3D frame data and using a two-stage training strategy. Scene-LLM's ability to understand egocentric and scene-level information enables it to handle scene changes and perform interactive planning in dynamic environments. The model's performance is validated through experiments on various tasks, demonstrating its effectiveness in 3D visual understanding and reasoning. Scene-LLM's contributions include introducing a 3D-VLM that connects 3D visual information with LLMs and achieving state-of-the-art results on 3D-VQA and interactive planning benchmarks. The model also proposes an effective 3D visual representation that supports state changes and can be easily integrated into LLMs. The model creates a large-scale dataset for 3D and text feature alignment, which includes 190k 3D-visual-language pairs from an egocentric viewpoint and about 500k pairs of scene-level data. Scene-LLM's performance is evaluated on various benchmarks, showing its effectiveness in 3D visual understanding and reasoning. The model's ability to understand egocentric and scene-level information enables it to handle scene changes and perform interactive