Research map / Vision-language and multimodal / Robot action and navigation
SOLAMI: Social Vision-Language-Action Modeling for Immersive Interaction with 3D Autonomous Characters
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Builds SOLAMI, an end-to-end social vision-language-action model that maps user speech and body motion to a 3D character's speech and motion tokens, trained on a synthesized multimodal interaction dataset SynMSI and deployed in a VR interface.
Problem
Equipping 3D autonomous characters with social intelligence to perceive, understand, and respond to users through speech and body language, where multimodal social interaction data is scarce and modular LLM-agent pipelines add latency and lose nuance.
Approach
Treats the digital character as a virtual humanoid robot and models it with an end-to-end VLA model: user speech and SMPL-X motion are discretized into tokens added to an LLM vocabulary, and the LLM generates response speech and motion tokens directly, trained with a three-stage pipeline on a synthesized dataset built from existing motion-text data.
Key results
- SOLAMI (full params) achieves FID 3.443 and PA-MPJPE 151.5 versus DLP (MotionGPT) at 4.254 and 165.053, with lowest latency of its variants and baselines (latency reported in Table 1).
- SOLAMI (full params) reaches the highest context relevance (3.824) and VC similarity (0.824) among compared methods on the SynMSI test split.
- In a 60-response VR user study, SOLAMI scored highest across motion coherence, motion interaction, speech consistency, and overall experience.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use end-to-end social VLA modeling instead of DLP's text-mediated LLM-Agent pipeline | SOLAMI · SynMSI test | FID: 4.254 -> 3.443 | Helped |
| use end-to-end social VLA modeling instead of DLP's text-mediated LLM-Agent pipeline | SOLAMI · SynMSI test | Inference Latency: 5.518 -> 2.639 | Helped |
| add the modality-alignment pre-training stage (ablation) | SOLAMI · SynMSI test | FID: 5.052 -> 3.443 | Helped |
| add the modality-alignment pre-training stage (ablation) | SOLAMI · SynMSI test | VC Similarity: 0.820 -> 0.824 | Helped |
| add the modality-alignment pre-training stage (ablation) | SOLAMI · SynMSI test | Character Consistency: 3.461 -> 3.824 | Helped |
| use full-parameter fine-tuning instead of LoRA fine-tuning (ablation) | SOLAMI · SynMSI test | FID: 15.729 -> 3.443 | Helped |
| use full-parameter fine-tuning instead of LoRA fine-tuning (ablation) | SOLAMI · SynMSI test | Diversity: 8.145 -> 8.853 | Helped |
| use full-parameter fine-tuning instead of LoRA fine-tuning (ablation) | SOLAMI · SynMSI test | VC Similarity: 0.770 -> 0.824 | Helped |
| use full-parameter fine-tuning instead of LoRA fine-tuning (ablation) | SOLAMI · SynMSI test | Context Relevance: 3.251 -> 3.634 | Helped |
| use full-parameter fine-tuning instead of LoRA fine-tuning (ablation) | SOLAMI · SynMSI test | Inference Latency: 2.710 -> 2.639 | No clear effect |
Surprising findings
- SOLAMI scores lower than LLM+Speech and AnyGPT on character consistency (2.639 vs 3.157 and 2.588 for AnyGPT; the DLP baseline scores 5.518), which the authors attribute to degradation of character knowledge in the LLM.
Limitations
- Character consistency is weaker than some baselines, likely due to degraded character knowledge after adding motion and speech modalities.
- Dataset is synthesized from existing motion datasets rather than collected real human-character interactions.
- Evaluation relies on GPT-4o as judge and a 60-response user study.
Experimental setup
Datasets: HumanML3D Inter-X DLP-MoCap CommonVoice AnyInstruct SynMSI
Benchmarks: SynMSI test split
Metrics: FID diversity PA-MPJPE angle error VC similarity context relevance character consistency inference latency user study Likert scores
Baselines: LLM+Speech (LLaMA2-7B-Chat) AnyGPT (fine-tune) DLP (MotionGPT) SOLAMI (w/o pretrain) SOLAMI (LoRA)
Builds on: AnyGPT MotionGPT SpeechTokenizer VQ-VAE DLP
Closest papers by meaning
- Digital Life Project: Autonomous 3D Characters with Social Intelligence (CVPR 2024 · similarity 0.87)
- 3D-VLA: A 3D Vision-Language-Action Generative World Model (ICML 2024 · similarity 0.85)
- SoMi-ToM: Evaluating Multi-Perspective Theory of Mind in Embodied Social Interactions (NeurIPS 2025 · similarity 0.85)
- The Language of Motion: Unifying Verbal and Non-verbal Language of 3D Human Motion (CVPR 2025 · similarity 0.85)
- Modeling Multimodal Social Interactions: New Challenges and Baselines with Densely Aligned Representations (CVPR 2024 · similarity 0.84)
- Human-Aware Vision-and-Language Navigation: Bridging Simulation to Reality with Dynamic Human Interactions (NeurIPS 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.