April 23-25, 2027 | Shenzhen, China
Special Session 2: Robot Skill Learning: From Demonstration to Generalization in the Perception–Decision–Action Loop
机器人技能学习:从示范到泛化的感知-决策-操作闭环
Introduction: Robot skill learning is entering a new phase — from single-task teaching toward generalization, transferability, and continual evolution. Vision-Language-Action (VLA) models have become general-purpose foundations, with Flow Matching and diffusion policies enabling high-frequency, multi-modal fine-grained action generation, while world models provide a generalization path via "imagined trial-and-error" for long-horizon tasks. In 2026, models such as Gemini Robotics 2 and Isaac GR00T demonstrate cross-embodiment deployment from a single checkpoint; Collective Experience Learning (CEL) validated a "the more you use, the stronger it gets" loop across 96 real robots; and Learning While Deploying (LWD) further boosts long-horizon success rates. Yet key bottlenecks remain: scarcity of high-quality real-robot data (available at hundreds-of-thousands to million-hour scale versus a demand of tens of millions of hours), the "morphology gap" in cross-embodiment transfer, sharp performance drops under distribution shift, and sample efficiency far below human levels. This Special Issue focuses on the core chain "from demonstration to generalization," bringing together the latest research on skill representation, data construction, policy learning, cross-embodiment transfer, and continual learning — advancing robots from "can demonstrate" to "can work."
Topics of interest for submission include, but are not limited to:
- Vision-Language-Action (VLA) Foundation Models and Action Representation. Including discrete action tokens, diffusion and flow-matching policies, action chunking, high-frequency real-time inference, and deployment-time compression.
- Hierarchical Skill Learning and Task-Planning Coordination. Including hierarchical architectures coupling high-level semantic planning with low-level closed-loop control, skill-library construction and composable invocation, and long-horizon task decomposition.
- World Models and "Imagination-Based" Skill Learning. Including world models for robotic manipulation, sample-efficient learning and policy optimization via imagined rollouts, and embodied memory mechanisms.
- Multi-Source Demonstration Data and Sim-to-Real Transfer. Including real-robot teleoperation and learning from human videos, synthetic simulation data, digital twins, domain randomization, and unified cross-modal data representations.
- Hybrid Paradigms of Imitation and Reinforcement Learning. Including offline RL, "imitation as the foundation, RL as the enhancer", learning from failures, reward design, and automatic curriculum generation.
- Few-Shot and Cross-Embodiment Skill Transfer. Including few-shot/zero-shot skill acquisition, cross-morphology and cross-scene generalization, meta-learning and test-time adaptation, and generalization benchmarks.
- Online Learning During Deployment and Collective Experience Sharing. Including Learning While Deploying, multi-robot experience aggregation and distribution, continual skill evolution, and mitigation of catastrophic forgetting.
- Multi-Modal Force-Tactile Sensing and Dexterous Manipulation. Including force/tactile sensing fusion, bimanual coordination, mobile manipulation, and skill learning for contact-rich and precision assembly tasks.
- Real-World Evaluation and Trustworthy Deployment. Including long-horizon reliability and safety verification, standardized benchmarks, deployment case studies in industrial/warehouse/home-service settings, and failure analysis.
Submission Guideline:
Please submit your manuscript via Online Submission System:
https://easychair.org/conferences/?conf=eecr2027
Please choose "Special Session 2"
Organizers:

Jing Luo, Wuhan University of Technology, China

Zhou ZHAO, Central China Normal University, China