POSTED: 02 Sep, 2024

Australian Cobotics Centre researchers have two papers accepted for publication at the upcoming IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) 2024 in Abu Dhabi. IROS is one of the largest and most important robotics research conferences in the world, attracting researchers, academics, and industry professionals from around the globe.

Postdoctoral Research Fellow, Dr Fouad Sukkar gave is a brief summary of two of the papers appearing at the conference in October this year.

Constrained Bootstrapped Learning for Few-Shot Robot Skill Adaptation, by Nadimul Haque, Fouad (Fred) Sukkar, Lukas Tanz, Marc Carmichael, Teresa Vidal Calleja, proposes a new method for teaching robot skills via demonstration. Often this is a cumbersome and time-consuming process since a human operator must provide a demonstration for every new task. Furthermore, there will inevitably be some discrepancies between how the demonstrator carries out the task versus the robot, for example, due to localisation errors, that need to be corrected for in order for the skill to be successfully transferred. This paper tackles these two problems by proposing a learning method that facilitates fast skill adoption to new tasks that have not been seen by the robot. We do so by training a reinforcement learning (RL) policy across a diverse set of scenarios in simulation offline and then use a sensor feedback mechanism to quickly refine the learnt policy to a new scenario with the real robot online. Importantly, to make offline learning tractable we utilise Hausdorff Approximation Planner (HAP) to constrain RL exploration to promising regions of the workspace. Experiments showcase our method achieving an average success rate of 90% across various complex manipulation tasks compared to state-of-the-art which only achieved 56%.

Coordinated Multi-arm 3D Printing using Reeb Decomposition, by Jayant Kumar , Fouad (Fred) Sukkar, Mickey Clemon, Ramgopal Mettu, proposes a framework for utilising multiple robot arms to collaboratively 3D print objects. For robots to do this efficiently and minimise downtime while printing, they must have the flexibility to work closely together in a shared workspace. However, this dramatically increases problem complexity since there is a need to coordinate the arms so they do not collide with each other or the partially printed object. This is in addition to the planning problem of effectively allocating parts of the object to each robot while respecting the physical dependencies of the print, for example an arm can’t start extruding a contour until all the contours below it are printed first. All these factors make effective coordination a very computationally hard problem and we show that with bad coordination you can end up with even worse utilisation than if a single arm had carried out the same print! In this work we address this by performing a Reeb decomposition of the object model which partitions the model into smaller, geometrically distinct components. This drastically reduces the search space over feasible toolpaths, thus allowing us to plan highly effective allocations to each arm using a tree search-based method. For producing fast collision avoiding motions we utilise Hausdorff Approximation Planner (HAP). Our experimental setup consists of two robot arms with pellet extruders mounted on their end effectors. We evaluate our framework on 14 different objects and show that our method achieves up to a mean utilisation improvement of 132% over benchmark methods.