Dr Dandan Zhang
ROBOTICS · AI · EDUCATION
IMPERIAL COLLEGE LONDON
Intelligence
How can learning systems turn perception into reliable physical action?
Six selected outcomes connect multimodal policies, feedback-driven planning, simulation and evaluation. Published and accepted conference work is identified separately from preprints.
Research overview · All publications
01 / ADAPTIVE POLICIES & FEEDBACK
ADM-DP · Adaptive vision, touch and coordination
ICRA · 2026 · Conference paper
Outcome. A diffusion policy adjusts the contribution of vision, touch and graph-based spatial information. Evaluation across seven manipulation tasks shows improved performance over the tested baselines, with tactile grasp refinement and shared spatial context supporting coordination.
Closed-loop multi-agent manipulation
Robotics: Science and Systems · 2026 · Conference paper
Outcome. Planning, manipulation and verification agents divide work and check physical execution. A verifier feeds corrections back into the plan; real-world experiments cover both shared and separate robot workspaces.
TacEvo · Discovering tactile learning architectures
IROS · 2026 · Accepted conference paper
Outcome. LLM-generated code changes are evaluated within a quality-diversity search. On ViTacTip tasks, the discovered models matched the expert force-prediction baseline and improved fine-grained grating classification in the reported post-search evaluation.
02 / SIMULATION, TRANSFER & CAPABILITY TESTING
Physics-informed learning for microscopy
ICRA · 2026 · Conference paper
Outcome. Wave-optics rendering and generative learning synthesise microscope images with depth-dependent diffraction and defocus. Pose estimators trained on these images approach real-data performance in the reported tests and generalise to unseen poses.
Interactive OT Gym · Learning microscale control
ICRA · 2025 · Conference paper and open platform
Outcome. An interactive optical-tweezers simulator combines physical modelling, reinforcement learning and haptic interfaces. It supports reproducible experiments with manual, autonomous and shared control before testing strategies on physical systems.
LLM4Robot-Bench · Testing physical interaction
Preprint · 2026 · Formerly PhysLadder; renamed 3 November 2026
Outcome. Sixteen matched simulation and real-world tasks test geometric alignment, sustained force regulation and coordinated contact reconfiguration. The hierarchical benchmark exposes where a generalist model’s local actions fail to produce reliable task completion.
Original ADM-DP figure and earlier studies in explainable hierarchical imitation learning, one-shot domain adaptation and service-robot learning follow.
 Add a TitleDescribe your image |  Add a TitleDescribe your image |  Add a TitleDescribe your image |
|---|---|---|
 Add a TitleDescribe your image |  Add a TitleDescribe your image |  Add a TitleDescribe your image |
Explainable Hierarchical Imitation Learning
Dandan Zhang; Qiang Li; Yu Zheng; Dongsheng Zhang; Zhengyou Zhang
To accurately pour drinks into various containers is an essential skill for service robots. However, drink pouring is a dynamic process and difficult to model.
Traditional deep imitation learning techniques for implementing autonomous robotic pouring have an inherent black-box effect and require a large amount of demonstration data for model training. To address these issues, an Explainable Hierarchical Imitation Learning (EHIL) method is proposed in this paper such that a robot can learn high-level general knowledge and execute low-level actions across multiple drink pouring scenarios. Moreover, with the EHIL method, a logical graph can be constructed for task execution, through which the decision-making process for action generation can be made explainable to users and the causes of failure can be traced out. Based on the logical graph, the framework is manipulable to achieve different targets while the adaptability to unseen scenarios can be achieved in an explainable manner.
A series of experiments have been conducted to verify the effectiveness of the proposed method. Results indicate that EHIL outperforms the traditional behavior cloning method in terms of success rate, adaptability, manipulability and explainability.
Full paper link:
https://ieeexplore.ieee.org/document/9667114
One-Shot Domain-Adaptive Imitation Learning
via Progressive Learning
Dandan Zhang; Wen Fan; John Llyod; Chenguang Yang;Nathan Lepora
Full paper link:
Traditional deep learning-based visual imitation learning techniques require a large amount of demonstration data for model training, and the pre-trained models are difficult to adapt to new scenarios. To address these limitations,
we propose a unified framework using a novel progressive learning approach comprised of three phases: i) a coarse learning phase for concept representation, ii) a fine learning phase for action generation, and iii) an imaginary learning phase for domain adaptation. Overall, this approach leads to a one-shot domain-adaptive imitation learning framework. We use robotic pouring task as an example to evaluate its effectiveness. Our results show that the method has several advantages over contemporary end-to-end imitation learning approaches, including an improved success rate for task execution and more efficient training for deep imitation learning. In addition, the generalizability to new domains is improved, as demonstrated here with novel background, target container and granule combinations. We believe that the proposed method can be broadly applicable to different industrial or domestic applications that involve deep imitation learning for robotic manipulation, where the target scenarios have high diversity while the human demonstration data is limited.
Robot Learning for Service Robots
Learning from Demonstration for Automatic Tea Preparation for Service Robots
AARON ASAMOAH; Virginia Ruiz Garate; Dandan Zhang