BayRnTune: Adaptive Bayesian Domain Randomization via Strategic Fine-tuning

Domain randomization (DR), which entails training a policy with randomized dynamics, has proven to be a simple yet effective algorithm for reducing the gap between simulation and the real world. However, DR often requires careful tuning of randomization parameters. Methods like Bayesian Domain Randomization (Bayesian DR) and Active Domain Randomization (Adaptive DR) address this issue by automating parameter range selection using real-world experience. While effective, these algorithms often require long computation time, as a new policy is trained from scratch every iteration. In this work, we propose Adaptive Bayesian Domain Randomization via Strategic Fine-tuning (BayRnTune), which inherits the spirit of BayRn but aims to significantly accelerate the learning processes by fine-tuning from previously learned policy. This idea leads to a critical question: which previous policy should we use as a prior during fine-tuning? We investigated four different fine-tuning strategies and compared them against baseline algorithms in five simulated environments, ranging from simple benchmark tasks to more complex legged robot environments. Our analysis demonstrates that our method yields better rewards in the same amount of timesteps compared to vanilla domain randomization or Bayesian DR.

Design of a Jumping Control Framework with Heuristic Landing for Bipedal Robots

Generating dynamic jumping motions on legged robots remains a challenging control problem as the full flight phase and large landing impact are expected. Compared to quadrupedal robots or other multi-legged robots, bipedal robots place higher requirements for the control strategy given a much smaller footprint. To solve this problem, a novel heuristic landing planner is proposed in this paper. With the momentum feedback during the flight phase, landing locations can be updated to minimize the influence of uncertainties from tracking errors or external disturbances when landing. To the best of our knowledge, this is the first approach to take advantage of the flight phase to reduce the impact of the jump landing which is implemented in the actual robot. By integrating it with a modified kino-dynamics motion planner with centroidal momentum and a low-level controller which explores the whole-body dynamics to hierarchically handle multiple tasks, a complete and versatile jumping control framework is designed in this paper. Extensive results of simulation and hardware jumping experiments on a miniature bipedal robot with proprioceptive actuation are provided to demonstrate that the proposed framework is able to achieve human-like efficient and robust jumping tasks, including directional jump, twisting jump, step jump, and somersaults.

Residual Physics Learning and System Identification for Sim-to-real Transfer of Policies on Buoyancy Assisted Legged Robots

The light and soft characteristics of Buoyancy Assisted Lightweight Legged Unit (BALLU) robots have a great potential to provide intrinsically safe interactions in environments involving humans, unlike many heavy and rigid robots. However, their unique and sensitive dynamics impose challenges to obtaining robust control policies in the real world. In this work, we demonstrate robust sim-to-real transfer of control policies on the BALLU robots via system identification and our novel residual physics learning method, Environment Mimic (EnvMimic). First, we model the nonlinear dynamics of the actuators by collecting hardware data and optimizing the simulation parameters. Rather than relying on standard supervised learning formulations, we utilize deep reinforcement learning to train an external force policy to match real-world trajectories, which enables us to model residual physics with greater fidelity. We analyze the improved simulation fidelity by comparing the simulation trajectories against the real-world ones. We finally demonstrate that the improved simulator allows us to learn better walking and turning policies that can be successfully deployed on the hardware of BALLU.

Benchmark Results for Bookshelf Organization Problem as Mixed Integer Nonlinear Program with Mode Switch and Collision Avoidance

Mixed integer convex and nonlinear programs, MICP and MINLP, are expressive but require long solving times. Recent work that combines data-driven methods on solver heuristics has shown potential to overcome this issue allowing for applications on larger scale practical problems. To solve mixed-integer bilinear programs online with data-driven methods, several formulations exist including mathematical programming with complementary constraints (MPCC), mixed-integer programming (MIP). In this work, we benchmark the performances of those data-driven schemes on a bookshelf organization problem that has discrete mode switch and collision avoidance constraints. The success rate, optimal cost and solving time are compared along with non-data-driven methods. Our proposed methods are demonstrated as a high level planner for a robotic arm for the bookshelf problem.

ReDUCE: Reformulation of Mixed Integer Programs using Data from Unsupervised Clusters for Learning Efficient Strategies

Mixed integer convex and nonlinear programs, MICP and MINLP, are expressive but require long solving times. Recent work that combines learning methods on solver heuristics has shown potential to overcome this issue allowing for applications on larger scale practical problems. Gathering sufficient training data to employ these methods still present a challenge since getting data from traditional solvers are slow and newer learning approaches still require large amounts of data. In order to scale up and make these hybrid learning approaches more manageable we propose ReDUCE, a method that exploits structure within small to medium size datasets. We also introduce the bookshelf organization problem as an MINLP as a way to measure performance of solvers with ReDUCE. Results show that existing algorithms with ReDUCE can solve this problem within a few seconds, a significant improvement over the original formulation. ReDUCE is demonstrated as a high level planner for a robotic arm for the bookshelf problem.

Deep Reinforcement Learning with Linear Quadratic Regulator Regions

Practitioners often rely on compute-intensive domain randomization to ensure reinforcement learning policies trained in simulation can robustly transfer to the real world. Due to unmodeled nonlinearities in the real system, however, even such simulated policies can still fail to perform stably enough to acquire experience in real environments. In this paper we propose a novel method that guarantees a stable region of attraction for the output of a policy trained in simulation, even for highly nonlinear systems. Our core technique is to use "bias-shifted" neural networks for constructing the controller and training the network in the simulator. The modified neural networks not only capture the nonlinearities of the system but also provably preserve linearity in a certain region of the state space and thus can be tuned to resemble a linear quadratic regulator that is known to be stable for the real system. We have tested our new method by transferring simulated policies for a swing-up inverted pendulum to real systems and demonstrated its efficacy.