Below are the Final Year Projects offered by the BionicDL lab this year. Students interested in any of these projects, please contact Prof. Song to express your interest of participation.

SLAM for an Amphibian Lobster Robot

Amphibian locomotion refers to a challenging scenario for robots moving in the field with the presence of both water and land, such as streams, wetland, mud, or even seabed locomotion besides normal land terrain. This project leverages a recent development at the BionicDL lab of a Lobster-inspired robot with overconstrained legs, aiming at the integration of a vision system for SLAM. While SLAM system for field robots has been well-researched in literature, it remains a challenging task for amphibious scenarios, where the presence of water at different levels may cause catastrophic challenges for legged robots commonly designed and trained for land locomotion. Further research in this area provides in-depth understanding of amphibious locmotion and its related SLAM system development, holding the potentials for autonomous navigation in challenging terrains.

The main tasks of this project involves three levels of development, including a water-proofed vision-based learning system, a remote joystick for field research with omni-directional control, and the SLAM algorithm development for amphibious envirionment. The vision system powered with learning computation should be capable of working in the field, even with the presence water. The remote joystick serves as a manual controller to work with the robot’s omni-directional locomotion, and finally the SLAM algorithms trained with amphibious scenarios for deployment and testing on the on-board computer.

The feasibility of this project is reflected in serveral aspects. The BionicDL lab has already designed and developed an prototype of the legged robot using overconstrained linkages as limbs, capable of omni-directional locomotion using only two motors per leg, providing the core platform of implementation and testing. The current computer used on this robot is based on the Nvidia Xavier platform, which is a high-performance, small-size, low-cost solution for field robots with various connections to external systems such as cameras, controllers, or run computational demanding tasks such as SLAM. Other components such as Intel RealSense D435i and Xbox One Joystick are also integrated for testing. The BionicDL lab is also equipped with a water tank of sufficient size to provide an controlled environment for testing, which is complemented by the vairous amphibious scenarios at SUSTech campus.

This project expects a workload of 20% on the water-proofed camera and computer, 20% on the remote controller integration, and 40% on the SLAM algorithm development, and a final 20% on any testing in simulation or field.

The project timeline is as follows:

Before Dec 15:
1. Literature review
2. Hardware pruchase & preparation
3. Benchmark dataset preparation
4. Experiment design
Before Jan 15
1. System integration with waterproofing and remote control
2. Simulation environment setup
3. Algorithm design
Before Apr 15
1. Full algorithm design and development
2. Dataset collection and simulated testing against benchmark
3. Preliminary system testing with the robot
Before May 15
1. Tentative field testing
2. Result summary and system review
3. Final report preparation and revision

Lobster-inspired Finger Surface Pattern Design

Contact plays a key role in securing a robust grasp towards dexterous object manipulation. Crustaceasn such as lobsters and crabs naturally habitat in extreme environment on the seabed with sands and floating particles underwate with low visibility, yet versatile in hunting and survival with only two rigidly-designed fingers per limb with only one degree-of-freedom, different from the multi-fingered design such as human. In this project, a bio-inspired approach is proposed to learn from the finger surface pattern design of the lobster and crab using soft, deformable materials. Using a parametric design method, a large collection of patterns of the finger surface are to be generated and benchmarked in terms of their friction performance. From a short-listed surface pattern, finger integration is to be implemented on a set of soft fingers with omni-direction adaptation during physical interaction. A test platform is to be developed to systematically experiment with the robotic system for robust grasping. Using the EGAD object set, the grasping performance of the finger surface patterns are to be benchmarked, which is to be finally tested with a learning-based grasping strategy to find the optimized design for learning-based two-finger grasping with robust performance.

The main tasks of this project involves a parametric design of finger surface pattern based on inspirations from the lobsters, design and development of a friction testing platform and an automated grasp testing platform, and a learning-based grasping system with the finger surface integrated. A collection of finger surfaces are to be designed and fabricated for testing. Extensive expeirments are to be carried out with an automated grasping platform and the EGAD object sets. And finally an integrated soft finger with an optimized finger surface for grasping applications.

The feasibility of this project are reflected in several aspects. Preliminary research on the lobster-inspired finger surface shows promising results of this approach. The BionicDL lab has been developing several research on the omni-direction soft finger with design iterations and system integration. In addition, the participant of this project can leverage the DeepClaw system developed at the BionicDL lab to perform systematic experiments.

The workload of this project is estimated with 30% on the parametric design analysis, 40% on developing and carrying out the systematic experiments, and 30% on the final grasping analysis with the learning system.

The project timeline is as follows:

Before Dec 15:
1. Literature review
2. Parametric finger surface design inspired by the lobsters
3. Experiment design
Before Jan 15
1. Experiment platform design and development
2. Finger surface design and fabrication
Before Apr 15
1. Full experiment implementation and data recording
2. Result analysis and review
Before May 15
1. Result summary and system review
2. Final report preparation and revision

Learning Overconstrained Locomotion over Amphibious Terrain

Walking over challenging terrains enables filed robots the capability to explore and navigate the unknowns. Recent progress such as the Spot Mini, ANYmal and Cheeta has shown various levels of legged locomotion, with a similar design of the leg with three degree-of-freedoms per leg. On the other hand, training robots to learn capable skills to walk over challenging terrains usually require a large amount simulation. In this project, a novel design of the legged robot is adopted using overconstrained linkages with dual actuations per leg, enabled the robot with omni-directional capability with fewer actuators, achieving a design with enhanced power-to-weight ratio for ground mobility. With this new design, this project aims at exploring the boundaries of its locomotive capability over amphibious terrain that is commonly presented and also challenging in the field.

The main tasks of this project involves setting up a simulated environment for the overconstrained legged robot with challenging terrains, developing an reinforcement learning algorithm for training omni-directional locomotion, and deployment in a physical robot with benchmarked results in amphibious terrains. Inspired by recent work published on the ANYmal, this project aims at developing a learning-based scheme using proprioceptive data to explore the potentials of overconstrained locomotion over challenging terrains, paving ways for field testing.

The feasibility of this project is reflected in several aspects. The BionicDL has been working on a preliminary prototype of the overconstrained robot with three design iterations, providing the necessary models and foundations for this project. In the meanwhile, fundamental research on the overconstrained locomotion of this robot prototypes has been carried out in two recent journal submissions with follow-up plans for long-term research, providing this project with the necessary support. Besides, another simplified model has also been developed with servo motors for rapid testing. The BionicDL lab is equipped with a high-performing, state-of-the-art workstation for demanding computation with learning-based research and simulation. Earlier research on this project shows a promising potential of this design for challenging terrains, which will be further explored in this project.

The workload of this project is estimated with 20% on setting up the various terrain scenarios for testing, 50% on developing the reinforcement learning algorithm, parameter tuning and system optimization, and 30% on the field testing with a robotic prototype.

The project timeline is as follows:

Before Dec 15:
1. Literature review
2. Review of the simulated environment
3. Review of the algorithm setup
Before Jan 15
1. Setting up the experiment experiment
2. Algorithm design for reinforcement learning
Before Apr 15
1. Systematic simulation over various terrain on overconstrained locomotion
2. Benchmarking results between parallel four-bar and overconstrained legs
Before May 15
1. Result summary and system review
2. Final report preparation and revision

Learning Overconstrained Locomotion using Volumetrically Enhanced Hybrid Actuator with Variable Stiffness

Robots built with soft materials are commonly designed with inspirations from nature. Building on the previous research at the BionicDL on the lobster-inspired robotic actuator, this project aims at exploring a design paradigm of soft actuators through volumetrically enhanced guidelines with kinematically defined models and much improved motion, force and torque outputs. By modulating the stiffness of the actuator body, the resultant design holds the potential for proprioceptive sensing with promising applications for industrial scenarios. In-depth modeling of the actuator’s kinematic and dynamic performance is to be carried out in this project, exploring the boundaries of the volumetrically enhanced soft actuator with rigid and soft components by learning overconstrained locomotion for amphibious Terrain.

The main tasks of this project involves in-depth theoretical modeling of the kinematic and dynamic behaviors of VEHA with double actuations, the design and fabrication of the rotary VEHA with double actuation and variable stiffness, and the investigation of a legged robot using VEHA with overconstrained locomotion through reinforcement learning.

The feasibility of this project involves a recent submission of earlier research on this design with single actuation. Earlier prototypes and preliminary experiment results showed promising outcomes with published results regarding soft actuators with rigid components, as inspired by the lobster. The BionicDL lab has equipped with the necessary experiment environment including pneumatic power source, fabrication tools and materials, related research experience, testing equipment and methods, and funding support from national science foundation, contributing to the feasibility of this project.

The workload of this project is estimated with 30% on theoretical modeling of VEHA with double actuation, 40% on the systematic experiments and integration of VEHA for stiffness control, and 30% on learning overconstrained locomotion through a legged robot built with VEHA.
The project timeline is as follows:

Before Dec 15:
1. Literature review
2. Theoretical modeling of the kinematics
3. Actuator design and experiment platform preparation
Before Jan 15
1. Stiffness control experiment
2. Legged robot design and integration
Before Apr 15
1. A full analysis of the actuator and the robot through experiment and learning
2. Result analysis and review
Before May 15
1. Result summary and system review
2. Final report preparation and revision

Ambient Intelligence for SuperLimb

SuperLimbs aims at exploring the brain functionality for extra limb motion assistance as a wearable device. However, for elderly users with reduced brain functionality, it becomes a dilemma between exploiting the brain function and detecting intention for machine-assisted decision and actuation. In this project, the goal is to systematically review the existing literature of superlimbs and further explore its boundaries for elderly users and the interactions, investigating the possibility of adopting ambient intelligences for assisted system control. Using tools such as depth camera, privacy-protected point cloud data could be collected to generate intention oriented data for analyzing the human intentions or status, providing computational support for robotic systems that works with the benefit of human. This project also leverages recent work with MIT on the development of a robotic cane as a superlimb.

The main tasks of this project involves an in-depth review of literature of superlimb for general use and elderly use, a comprehensive review of literature of ambient intelligence and its potential integration with robotic systems such as superlimb for human assistance, and experimental validation and optimization based on the existing supercane system. This project is highly feasible as its built upon previous published work on the supercane in collaboration with MIT, with two iterations of the preliminary prototype developed.

The workload of this project is estimated with 40% on the review of superlimb, 40% on the review of ambient intelligence, and 30% on the system integration with the supercane as an exemplary research.

The project timeline is as follows:

Before Dec 15:
1. Literature review on superlimb
Before Jan 15
1. Literature review on ambient intelligence
Before Apr 15
1. At least three in-depth iterations of the review
2. At least one in-depth iteration of the supercane system
Before May 15
1. Result summary and system review
2. Final report preparation and revision

1 SLAM for an Amphibian Lobster Robot
2 Lobster-inspired Finger Surface Pattern Design
3 Learning Overconstrained Locomotion over Amphibious Terrain
4 Learning Overconstrained Locomotion using Volumetrically Enhanced Hybrid Actuator with Variable Stiffness
5 Ambient Intelligence for SuperLimb