Apan Dastider, PhD

“Every brilliant experiment, like every great work of art, starts with an act of imagination.” – Jonah Lehrer 

Projects

RESEARCH PROJECTS

• APEX : Collision free dual arm manipulation by Generative Artificial Intelligence Models [Demo ]

o Simplifying real-world complex dual-arm manipulation tasks by abstracting them simple vector alignment tasks between different objects
o Utilizing a latent stable diffusion models for creating various dual arm manipulation trajectories
o Integrating obstacle-guidance technique in the diffusion models for collision free trajectories with environmental obstacles in the workspace

• Online Adjustment : Sub-optimal Trajectory Refinement Through Linearized Value Function Computation [ Demo ]

o Achieving dynamic adaptability in non-stationary environment through linear computation efficient value function computation

o Refinement of Sub-optimal action sequences from DDP or iLQR for evolving target objective function through linearized feedback of adjusted exponentiated value function

o Computation of distribution shift through KL divergence as desirability function and Optimal Action adjustment through feeding differentiable desirability function values computed in linearized computation efficient mechanisms

• Robotic Motion Planning through Low-dimensional Manifold Building [ Demo ]  
  1. Amorphous Obstacle Avoidance:
      o Proposed a obstacle avoidance algorithm where the 3D geometry of obstacles can reshape at any instance
      o Developed a geometrical manifold learning algorithm for introducing densely connected 2-D graph of low-dimensional points
      o Smooth Motion Planning over the connected graph and deep-autoencoding for two-way mapping between high-dimensional robotic state and low-dimensional manifold representation
  2. Reactive Whole-Body Obstacle Avoidance for Collision-Free Human-Robot Interaction [ Demo ]
      o Topological low-dimensional Manifold Building with dual arm joint pose high dimensional state space
      o Manipulation over the connected collision-free points for safer robotic motion planning and location reaching smoothly
      o Optimal reactive sensor placement on robotic manipulator's mechanical structure for the best possible occlusion detection 

3. Multi-Robot Skill Learning from a unified manifold representation in real-time  [Demo]
    o Learning multiple robotic skills -- obstacle avoidance and interception of flying targets, adaptively in real-time through a 2D graph traversing
    o Leveraging an innovative diffusion map based manifold learning algorithm for transforming high-dimensional robotic state-space to low-dimensional latent space representation
    o Extended Kalman Filter (EKF) for target location estimation for moving target for accurate interception 

• Non-parametric Deep Reinforcement Learning
  1. Strategic Retreats for Learning sequence of Optimal Control Policies [ Website ]
      o Developed non-parametric representation of stochastic policy gradients and exploited the blessings of vector-space properties of RKHS for incorporating a check-point detection method
      o Proposed a Gradient Similarity check method for pre-mature action abortion and only executing the optimal part of the action sequence
      o Generating sequence of control policies for handling bounded non-stationarity and maximizing utility in RL framework
  2. RKHS embeddings as priors in Bayesian Policy Gradients [ Demo ]
      o Enabling a robotic agent to incrementally bias its current learning process through utilizing prior relevant interactions by following Bayesian Exploration Strategy
      o Representation of Prior distributions in a tractable and linearly complex kernel space, RKHS
      o Using a dot product based similarity metric to prioritize past-task transitions to jump-start current learning process 

ACADEMIC PROJECTS

• Learning Safe locomotion by bio-robots with deep Reinforcement Learning [Course : Robots, Agents and Humans]

• Fusion of Classical Control Theories with Reinforcement Learning : Address the 'sim-to-real' gap through introducing MRAC system in real-world agents with learned knowledge from simulated platforms n deep Reinforcement Learning framework. [Course : Adaptive Control]

• Learn inverse kinematics of a redundant robotic manipulators through using Deep Neural Networks [Course : Machine Learning]

UNDERGRADUATE INTERN PROJECTS 

(Surge Engineering, Dhaka, Bangladesh)

• Development of Software Tools, Control Circuitry and Embedded Systems for a locally developed low-cost thumbstick-controlled motorized wheelchair. [ Demo ] 

Publications

ARXIV  PRE-PRINT : 

• Reactive Whole-Body Obstacle Avoidance for Collision-Free Human-Robot Interaction with Topological Manifold Learning.  Apan Dastider, Mingjie Lin. arxiv preprint | website 

• Cross-Embodiment Robotic Manipulation Synthesis via Guided Demonstrations through CycleVAE and Human Behavior Transformer. Apan Dastider, Hao Fang, Mingjie Lin. Under Review in 2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). arxiv preprint

ACCEPTED : 

• Unified Control Framework for Real-Time Interception and Obstacle Avoidance of Fast-Moving Objects with Diffusion Variational Autoencoder. Apan Dastider, Hao Fang, Mingjie Lin. 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), UAE, 2024. arxiv preprint | website

• APEX: Ambidextrous Dual-Arm Robotic Manipulation Using Collision-Free Generative Diffusion Models. Apan Dastider, Hao Fang, Mingjie Lin. 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), UAE, 2024. website 

• RETRO: Reactive Trajectory Optimization for Real-Time Robot Motion Planning in Dynamic Environments. Apan Dastider, Hao Fang, Mingjie Lin. Accepted in 2024 IEEE International Conference on Robotics and Automation (ICRA) website |  arxiv preprint

• Dynamically Avoiding Amorphous Obstacles with Topological Manifold Learning and Deep Autoencoding. Apan Dastider, Mingjie Lin. (accepted in IROS 2023) demo 

• Non-Parametric Stochastic Policy Gradient with Strategic Retreat for Non-Stationary Environment. Apan Dastider, Mingjie Lin. Accepted in 2022 IEEE 18th International Conference on Automation Science and Engineering (CASE), Mexico City, Mexico. (in press)  arxiv preprint | website

•  Learning Adaptive Control in Dynamic Environments using Reproducing Kernel Priors with Bayesian Policy Gradients. A. Dastider, S.J.A Raza, M. Lin. Proceedings of the 2021 37th Annual ACM/SIGAPP Symposium on Applied Computing. Association for Computing Machinery, New York, NY, USA. (In Press) website/video demo 

• Safe Locomotion Within Confined Workspace using Deep Reinforcement Learning. A. Dastider, S.J.A Raza, M. Lin. 2021 Fifth IEEE International Conference on Robotic Computing (IRC), Taiwan. paper 

• Survivable Robotic Control through Guided Bayesian Policy Search with Deep Reinforcement Learning. S.J.A Raza, A. Dastider, M. Lin. 2021 IEEE 17th International Conference on Automation Science and Engineering (CASE), Lyon, France. paper | website/video demo

• Survivable Hyper-Redundant Robotic Arm with Bayesian Policy Morphing. S.J.A Raza, A. Dastider, M. Lin. 2020 16th IEEE-RAS International Conference on Automation Science and Engineering (CASE), Hong Kong. paper | website

• Developmentally Synthesizing Earthworm-Like Locomotion Gaits with Bayesian-Augmented DDPG. S.J.A Raza, A. Dastider, M. Lin. 2020 16th IEEE-RAS International Conference on Automation Science and Engineering (CASE), Hong Kong. paper | website