I am a postdoctoral researcher at the Aalto Robot Learning Lab, Aalto University, working with Prof. Joni Pajarinen. I study how learning agents coordinate — with each other and with the physical world: multi-agent reinforcement learning, object-centric representations of video, and transfer from human video to robot embodiments.
I received my Ph.D. in Computer Science and Technology from the University of Electronic Science and Technology of China (2024) and my B.S. from North China Electric Power University (2019). During my Ph.D. I spent a year at Aalto as a CSC-funded visiting researcher.
Under standard generative objectives, sparse multi-body motion data cannot pin down which source-to-target map a cross-skeleton retargeting model learns, so the paper proposes Source-Instance Fidelity (SIF) to test whether generated motion actually follows its source clip.
A simple masking strategy for diffusion training that targets spatially inconsistent generation: masking up to 98% of pixels keeps FID competitive, avoids training instability on small datasets and reduces memorization.
Learned temporal predictors in video object-centric models mostly just keep slot indices stable, so seeding slots from saliency peaks in frozen DINOv2 features and Hungarian-matching them across frames does the same job with no learned temporal parameters.
Implicit Cycle Consistency moves the cycle-consistency constraint from slot space to the reconstruction manifold, which avoids the feature collapse caused by explicit cycle consistency on slots and outperforms those baselines on video object-centric benchmarks.
Each agent plans with a vision-language model in a closed loop, writes and reuses code skills, and shares observations through multi-hop communication; evaluated on SMACv2.
Turns an egocentric human manipulation video into a robot demonstration video by separating task and embodiment codes that condition a frozen video diffusion model.
A controlled study showing that neither a learned temporal predictor nor a frozen backbone is always necessary for video object-centric learning: what matters is how object identity is carried across frames, and whether an adapted encoder stays tied to a frozen pretrained feature target.
Uses the number of agents as a curriculum variable and advances the curriculum with a TD-error-based learning-progress measure, outperforming state-of-the-art baselines on three sparse-reward MARL benchmarks.
A centralized joint policy correlates agents' actions, while a mode-consistency constraint keeps it executable by decentralized, partially observing agents.
Adaptive graph attention networks with interactive learning (AGAT-IL) for clustering the nodes of attributed graphs.
NcHo-AC frames adaptive traffic signal control as a competitive-cooperative game and combines neighborhood coordination with a mean-field approximation, so multi-agent actor-critic learning scales to large traffic networks.
Clipping the advantage to remove negative values gives optimistic multi-agent policy-gradient updates that counter relative overgeneralization, outperforming strong baselines on 13 of 19 tasks and matching them on the rest.
Later agents in an auto-regressive multi-agent policy send gradients back to earlier agents through their actions.
A hierarchical fusion framework for generating responses in multimodal dialogue.
Independent learners infer teammates' intentions from observed actions and choose whom to coordinate with — no centralized critic, no communication.
A parallel multi-agent deep RL framework that performs all three stages of NFV resource allocation jointly and online by generating an embedding subgraph per request.
