Robot Learning and Representation Learning (RL2)
AI Institute / School of Computer Science
Shanghai Jiao Tong University (SJTU)
Shanghai Innovation Institute (SII)
（上海交通大学 / 上海创智学院）
At RL2 Lab at SJTU/SII, we are always looking for passionate innovators to push the boundaries of AI and robotics. Our passion lies in representation of the real world, physical intelligence for humanoid robots and building empathetic systems that understand human emotion. We provide the platform, computational resources, and mentorship to turn ideas into impactful research. We warmly welcome highly motivated PhD, Master's, and undergraduate students to join our interdisciplinary team.
在上海交通大学/上海创智学院（SJTU/SII）的 RL² 实验室，我们始终在寻找充满激情的创新者，共同拓宽人工智能与机器人的边界。我们的研究热忱在于真实世界的表征、人形机器人的物理智能，以及构建能够理解人类情感的共情系统。我们提供硬件平台、计算资源和导师指导，将创新想法转化为具有影响力的科研成果。热烈欢迎具有强烈内驱力的博士生、硕士生及本科生加入我们的交叉学科团队。
鲁棒与安全的足式智能：聚焦人形、四足及六足机器人在复杂非结构化环境中的高动态感知、控制与安全部署。我们的研究致力于突破机器人在负载突变、硬件退化以及未知物理扰动等极端工况下的运动极限，赋予其在复杂地形中稳定、可靠的作业能力。
真实世界中的机器人技能学习：结合强化学习与表征学习，突破高效的 Sim-to-Real 迁移瓶颈。通过从人类示范、视频及轨迹数据中提取先验知识，赋予机器人平滑行走、复杂节律运动、精准全身模仿以及跨越极高难度地形的泛化与自适应能力。
视觉-语言-动作与具身导航：面向开放世界环境，构建从大尺度场景理解到复杂任务执行的端到端框架。将多模态视觉与自然语言指令深度融合，实现高精度的视觉语言导航与长程任务下的精准动作生成，打通从语义理解到物理交互的闭环。
情感化具身交互与共情系统：探索人机共融的新范式，打造具备情感感知与情绪共鸣的具身机器人。融合多模态情绪识别、共情行为决策与细腻的情感姿态表达技术，为机器人赋予理解和反馈人类情感的能力，实现有温暖且富有同理心的陪伴。
Representative works grouped by research direction. For the full list, see Publications.
Active Exploration and Online Perception of Terrain Physics with Legged Robots
RA-L 2025
Active exploration strategy coupled with online terrain-physics estimation for safe legged locomotion on varied surfaces.
Contrastive Forward Prediction Reinforcement Learning for Adaptive Fault-Tolerant Legged Robots
CoRL 2025
A cerebellum-inspired dual-pathway architecture for legged robots that adapts gait and tolerates joint failures in real time.
RA-L 2024
A novel policy parameterization based on the Dirichlet distribution that guarantees zero constraint violations for safe sim-to-real transfer.
Select before Act: Spatially Decoupled Action Repetition for Continuous Control
ICLR 2025
Spatially selective action repetition that improves sample efficiency and robustness in continuous locomotion control tasks.
HiWET: Hierarchical World-Frame End-Effector Tracking for Long-Horizon Humanoid Loco-Manipulation
RSS 2026
Hierarchical framework for long-horizon humanoid loco-manipulation; world-frame end-effector tracking decouples locomotion and manipulation, enabling robust whole-body control for complex real-world tasks.
A2CF: Learning Motion Skills with Adaptive Assistive Curriculum Force in Humanoid Robots
ICRA 2026
Assistive-force curriculum co-training with the robot's motion policy: applies assistive forces during early learning and gradually withdraws them, enabling stable walking, dancing, and backflip skills.
Keep on Going: Learning Robust Humanoid Motion Skills via Selective Adversarial Training
AAAI 2026
Non-zero-sum adversarial training with a budget-constrained attack policy; enables humanoid robots to traverse complex terrain and stairs in the real world.
Coordinated Humanoid Robot Locomotion with Symmetry Equivariant Reinforcement Learning Policy
AAAI 2026
Symmetry-equivariant actor and symmetry-invariant critic improve multi-directional tracking success; outperforms DreamWaQ by 1.3× on challenging scenarios.
Robust Locomotion Policy with Adaptive Lipschitz Constraint for Legged Robots
RA-L 2024
Lipschitz-constrained reinforcement learning that produces smoother and more disturbance-robust locomotion policies.
Global-Local Attention Decomposition for Terrain Encoding in Humanoid Perceptive Locomotion
Under Review
Cross-attention over height-map features with proprioceptive queries; foothold-point features dynamically focus the policy on traversable regions, deployed on G1 with onboard LiDAR.
GATER: Learning Grasp-Action-Target Embeddings and Relations for Task-Specific Grasping
RA-L 2022
Jointly learns grasp-action-target embeddings and their relations to enable task-specific grasping, bridging the gap between object affordances and downstream manipulation goals.
FocusNav: Spatial Selective Attention with Waypoint Guidance for Humanoid Local Navigation
Under Review
Spatial selective attention with waypoint guidance improves humanoid local navigation by focusing on task-relevant regions and planning robust traversal paths in cluttered environments.
Hierarchical Humanoid Manipulation with Physically Grounded Motion Intention Representations
Under Review, RA-L 2026
Physically grounded motion intention representations enable hierarchical decomposition of complex humanoid manipulation tasks for robust real-world execution.
Multi-Modal Hierarchical Empathetic Framework for Social Robots With Affective Body Control
TAFFC 2024
Multimodal affective understanding and hierarchical empathetic body-behavior generation for social robots in real-world HRI.
DanceHAT: Generate Stable Dances for Humanoid Robots with Adversarial Training
ICRA 2022
Adversarial training framework for generating physically stable and expressive dance motions for humanoid social robots from human demonstrations.
实验室长期招收计算机、自动化等方向硕博学生与博士后，欢迎有志于科研的本科生加入。
Office: Artificial Intelligence Institute 5-506 | Email: yuegao@sjtu.edu.cn