🤖 Robotics & Embodied AI

💬 ACL2025 · 7 paper notes

📌 Same area in other venues: 📷 CVPR2026 (146) · 🔬 ICLR2026 (162) · 💬 ACL2026 (11) · 🧪 ICML2026 (53) · 🤖 AAAI2026 (30) · 🧠 NeurIPS2025 (75)

🔥 Top topics: Robotics ×5 · Sentiment Analysis ×2

CHEER-Ekman: Fine-grained Embodied Emotion Classification

This paper proposes the CHEER-Ekman dataset, extending the binary embodied emotion annotations of the CHEER dataset into Ekman's six basic emotions. It employs an LLM-based automatic Best-Worst Scaling (BWS) technique to achieve fine-grained emotion classification without task-specific training, outperforming supervised BERT.

Rolling the DICE on Idiomaticity: How LLMs Fail to Grasp Context

This work proposes the DICE dataset (2066 sentences, 402 idioms) to reveal systematic flaws in LLMs when contextual understanding is required to disambiguate idioms (literal vs. figurative meanings), achieved through highly controlled contrastive evaluation with identical idiom forms.

Do Emotions Really Affect Argument Convincingness? A Dynamic Approach with LLM-based Manipulation Checks

This paper proposes a dynamic framework inspired by psychological manipulation checks, utilizing LLMs to modulate the emotional intensity of arguments and systematically investigate the causal impact of emotion on argument convincingness. The findings reveal that in more than half of the cases, human judgments of convincingness are unaffected by emotional changes; when emotion does have an effect, it is more likely to enhance rather than diminish convincingness.

DRAE: Dynamic Retrieval-Augmented Expert Networks for Lifelong Learning and Task Adaptation in Robotics

This paper proposes the DRAE framework, which integrates dynamic MoE routing, parametric RAG (P-RAG), a three-layer cognitive control architecture (ReflexNet-SchemaPlanner-HyperOptima), and DPMM lifelong knowledge retention. It achieves an average success rate of 82.5% on robotic manipulation and autonomous driving tasks, effectively mitigating catastrophic forgetting.

Task-aware MoILE: Hierarchical-Task-Aware Multi-modal Mixture of Incremental LoRA Experts for Embodied Continual Learning

This paper proposes a Hierarchical Embodied Continual Learning (HEC) setting, which divides agent learning into high-level instructions and low-level actions. It designs the Task-aware MoILE method—which automatically identifies tasks through cross-modal clustering, selects LoRA experts using dual routers, and retains past knowledge via SVD orthogonal training. It reduces the forgetting rate to 3.37% across 5 incremental learning scenarios (vs. 7.44% for the previous SOTA).

SELF-PERCEPT: Introspection Improves LLMs' Detection of Multi-Person Mental Manipulation in Conversations

This paper proposes the SELF-PERCEPT two-stage prompting framework, drawing on psychological Self-Perception Theory. It guides LLMs to first observe the behavioral cues of conversational participants before inferring their internal attitudes, significantly improving the detection of mental manipulation in multi-person, multi-turn dialogues.

Vulnerability of LLMs to Vertically Aligned Text Manipulations

This paper systematically reveals the severe vulnerability of LLMs to vertically aligned text inputs: vertically aligning only a small number of keywords can lead to a drop of 25-45 percentage points in text classification accuracy. While CoT reasoning fails to mitigate this issue, a well-designed few-shot learning paradigm can effectively recover performance.