Wenhao Yu

J. Taery Kim, Wenhao Yu, Yash Kothari et al. · gatech

This paper explores the principles for transforming a quadrupedal robot into a guide robot for individuals with visual impairments. A guide robot has great potential to resolve the limited availability of guide animals that are accessible to only two to three percent of the potential blind or visually impaired (BVI) users. To build a successful guide robot, our paper explores three key topics: (1) formalizing the navigation mechanism of a guide dog and a human, (2) developing a data-driven model of their interaction, and (3) improving user safety. First, we formalize the wayfinding task of the human-guide robot team using Markov Decision Processes based on the literature and interviews. Then we collect real human-robot interaction data from three visually impaired and six sighted people and develop an interaction model called the ``Delayed Harness'' to effectively simulate the navigation behaviors of the team. Additionally, we introduce an action shielding mechanism to enhance user safety by predicting and filtering out dangerous actions. We evaluate the developed interaction model and the safety mechanism in simulation, which greatly reduce the prediction errors and the number of collisions, respectively. We also demonstrate the integrated system on a quadrupedal robot with a rigid harness, by guiding users over $100+$~m trajectories.

Yihang Yao, Zhepeng Cen, Wenhao Ding et al. · cmu

Offline safe reinforcement learning (RL) aims to train a policy that satisfies constraints using a pre-collected dataset. Most current methods struggle with the mismatch between imperfect demonstrations and the desired safe and rewarding performance. In this paper, we introduce OASIS (cOnditionAl diStributIon Shaping), a new paradigm in offline safe RL designed to overcome these critical limitations. OASIS utilizes a conditional diffusion model to synthesize offline datasets, thus shaping the data distribution toward a beneficial target domain. Our approach makes compliance with safety constraints through effective data utilization and regularization techniques to benefit offline safe RL training. Comprehensive evaluations on public benchmarks and varying datasets showcase OASIS's superiority in benefiting offline safe RL agents to achieve high-reward behavior while satisfying the safety constraints, outperforming established baselines. Furthermore, OASIS exhibits high data efficiency and robustness, making it suitable for real-world applications, particularly in tasks where safety is imperative and high-quality demonstrations are scarce.

Yuming Feng, Chuye Hong, Yaru Niu et al.

Recently, quadrupedal locomotion has achieved significant success, but their manipulation capabilities, particularly in handling large objects, remain limited, restricting their usefulness in demanding real-world applications such as search and rescue, construction, industrial automation, and room organization. This paper tackles the task of obstacle-aware, long-horizon pushing by multiple quadrupedal robots. We propose a hierarchical multi-agent reinforcement learning framework with three levels of control. The high-level controller integrates an RRT planner and a centralized adaptive policy to generate subgoals, while the mid-level controller uses a decentralized goal-conditioned policy to guide the robots toward these sub-goals. A pre-trained low-level locomotion policy executes the movement commands. We evaluate our method against several baselines in simulation, demonstrating significant improvements over baseline approaches, with 36.0% higher success rates and 24.5% reduction in completion time than the best baseline. Our framework successfully enables long-horizon, obstacle-aware manipulation tasks like Push-Cuboid and Push-T on Go1 robots in the real world.

Yucheng Shi, Wenhao Yu, Zaitang Li et al.

Recently, there has been a surge of vision-based GUI agents designed to automate everyday mobile and web tasks. These agents interpret raw GUI screenshots and autonomously decide where to click, scroll, or type, which bypasses handcrafted rules and app-specific APIs. However, most existing methods trained GUI agent in the offline environment using pre-collected trajectories. This approach limits scalability, causes overfitting to specific UI templates, and leads to brittle policies when faced with unseen environment. We present MobileGUI-RL, a scalable framework that trains GUI agent in online environment. MobileGUI-RL contains two key components. It (i) synthesizes a curriculum of learnable tasks through self-exploration and filtering, and (ii) adapts GRPO to GUI navigation with trajectory-aware advantages and composite rewards that balance task success and execution efficiency. Experiments on three online mobile-agent benchmarks show consistent gains, validating the effectiveness of our approach.

Yunbo Zhang, Wenhao Yu, C. Karen Liu et al.

We present a method of training character manipulation of amorphous materials such as those often used in cooking. Common examples of amorphous materials include granular materials (salt, uncooked rice), fluids (honey), and visco-plastic materials (sticky rice, softened butter). A typical task is to spread a given material out across a flat surface using a tool such as a scraper or knife. We use reinforcement learning to train our controllers to manipulate materials in various ways. The training is performed in a physics simulator that uses position-based dynamics of particles to simulate the materials to be manipulated. The neural network control policy is given observations of the material (e.g. a low-resolution density map), and the policy outputs actions such as rotating and translating the knife. We demonstrate policies that have been successfully trained to carry out the following tasks: spreading, gathering, and flipping. We produce a final animation by using inverse kinematics to guide a character's arm and hand to match the motion of the manipulation tool such as a knife or a frying pan.

Wenhao Yu, C. Karen Liu, Greg Turk

Being able to transfer existing skills to new situations is a key capability when training robots to operate in unpredictable real-world environments. A successful transfer algorithm should not only minimize the number of samples that the robot needs to collect in the new environment, but also prevent the robot from damaging itself or the surrounding environment during the transfer process. In this work, we introduce a policy transfer algorithm for adapting robot motor skills to novel scenarios while minimizing serious failures. Our algorithm trains two control policies in the training environment: a task policy that is optimized to complete the task of interest, and a protective policy that is dedicated to keep the robot from unsafe events (e.g. falling to the ground). To decide which policy to use during execution, we learn a safety estimator model in the training environment that estimates a continuous safety level of the robot. When used with a set of thresholds, the safety estimator becomes a classifier for switching between the protective policy and the task policy. We evaluate our approach on four simulated robot locomotion problems and a 2D navigation problem and show that our method can achieve successful transfer to notably different environments while taking the robot's safety into consideration.

Ioannis Exarchos, Yifeng Jiang, Wenhao Yu et al.

Transferring reinforcement learning policies trained in physics simulation to the real hardware remains a challenge, known as the "sim-to-real" gap. Domain randomization is a simple yet effective technique to address dynamics discrepancies across source and target domains, but its success generally depends on heuristics and trial-and-error. In this work we investigate the impact of randomized parameter selection on policy transferability across different types of domain discrepancies. Contrary to common practice in which kinematic parameters are carefully measured while dynamic parameters are randomized, we found that virtually randomizing kinematic parameters (e.g., link lengths) during training in simulation generally outperforms dynamic randomization. Based on this finding, we introduce a new domain adaptation algorithm that utilizes simulated kinematic parameters variation. Our algorithm, Multi-Policy Bayesian Optimization, trains an ensemble of universal policies conditioned on virtual kinematic parameters and efficiently adapts to the target environment using a limited number of target domain rollouts. We showcase our findings on a simulated quadruped robot in five different target environments covering different aspects of domain discrepancies.

Wenhao Yu, Jie Tan, Yunfei Bai et al.

The ability to walk in new scenarios is a key milestone on the path toward real-world applications of legged robots. In this work, we introduce Meta Strategy Optimization, a meta-learning algorithm for training policies with latent variable inputs that can quickly adapt to new scenarios with a handful of trials in the target environment. The key idea behind MSO is to expose the same adaptation process, Strategy Optimization (SO), to both the training and testing phases. This allows MSO to effectively learn locomotion skills as well as a latent space that is suitable for fast adaptation. We evaluate our method on a real quadruped robot and demonstrate successful adaptation in various scenarios, including sim-to-real transfer, walking with a weakened motor, or climbing up a slope. Furthermore, we quantitatively analyze the generalization capability of the trained policy in simulated environments. Both real and simulated experiments show that our method outperforms previous methods in adaptation to novel tasks.

Yunbo Zhang, Wenhao Yu, Greg Turk

In this work, we present a reinforcement learning algorithm that can find a variety of policies (novel policies) for a task that is given by a task reward function. Our method does this by creating a second reward function that recognizes previously seen state sequences and rewards those by novelty, which is measured using autoencoders that have been trained on state sequences from previously discovered policies. We present a two-objective update technique for policy gradient algorithms in which each update of the policy is a compromise between improving the task reward and improving the novelty reward. Using this method, we end up with a collection of policies that solves a given task as well as carrying out action sequences that are distinct from one another. We demonstrate this method on maze navigation tasks, a reaching task for a simulated robot arm, and a locomotion task for a hopper. We also demonstrate the effectiveness of our approach on deceptive tasks in which policy gradient methods often get stuck.

Wenhao Yu, Visak CV Kumar, Greg Turk et al.

We present a new approach for transfer of dynamic robot control policies such as biped locomotion from simulation to real hardware. Key to our approach is to perform system identification of the model parameters μ of the hardware (e.g. friction, center-of-mass) in two distinct stages, before policy learning (pre-sysID) and after policy learning (post-sysID). Pre-sysID begins by collecting trajectories from the physical hardware based on a set of generic motion sequences. Because the trajectories may not be related to the task of interest, presysID does not attempt to accurately identify the true value of μ, but only to approximate the range of μ to guide the policy learning. Next, a Projected Universal Policy (PUP) is created by simultaneously training a network that projects μ to a low-dimensional latent variable η and a family of policies that are conditioned on η. The second round of system identification (post-sysID) is then carried out by deploying the PUP on the robot hardware using task-relevant trajectories. We use Bayesian Optimization to determine the values for η that optimizes the performance of PUP on the real hardware. We have used this approach to create three successful biped locomotion controllers (walk forward, walk backwards, walk sideways) on the Darwin OP2 robot.

Wenhao Yu, C. Karen Liu, Greg Turk

Computer simulation provides an automatic and safe way for training robotic control policies to achieve complex tasks such as locomotion. However, a policy trained in simulation usually does not transfer directly to the real hardware due to the differences between the two environments. Transfer learning using domain randomization is a promising approach, but it usually assumes that the target environment is close to the distribution of the training environments, thus relying heavily on accurate system identification. In this paper, we present a different approach that leverages domain randomization for transferring control policies to unknown environments. The key idea that, instead of learning a single policy in the simulation, we simultaneously learn a family of policies that exhibit different behaviors. When tested in the target environment, we directly search for the best policy in the family based on the task performance, without the need to identify the dynamic parameters. We evaluate our method on five simulated robotic control problems with different discrepancies in the training and testing environment and demonstrate that our method can overcome larger modeling errors compared to training a robust policy or an adaptive policy.

Wenhao Yu, C. Karen Liu, Greg Turk

We present a method for efficient learning of control policies for multiple related robotic motor skills. Our approach consists of two stages, joint training and specialization training. During the joint training stage, a neural network policy is trained with minimal information to disambiguate the motor skills. This forces the policy to learn a common representation of the different tasks. Then, during the specialization training stage we selectively split the weights of the policy based on a per-weight metric that measures the disagreement among the multiple tasks. By splitting part of the control policy, it can be further trained to specialize to each task. To update the control policy during learning, we use Trust Region Policy Optimization with Generalized Advantage Function (TRPOGAE). We propose a modification to the gradient update stage of TRPO to better accommodate multi-task learning scenarios. We evaluate our approach on three continuous motor skill learning problems in simulation: 1) a locomotion task where three single legged robots with considerable difference in shape and size are trained to hop forward, 2) a manipulation task where three robot manipulators with different sizes and joint types are trained to reach different locations in 3D space, and 3) locomotion of a two-legged robot, whose range of motion of one leg is constrained in different ways. We compare our training method to three baselines. The first baseline uses only joint training for the policy, the second trains independent policies for each task, and the last randomly selects weights to split. We show that our approach learns more efficiently than each of the baseline methods.

Alexander Clegg, Wenhao Yu, Jie Tan et al.

We investigate robotic assistants for dressing that can anticipate the motion of the person who is being helped. To this end, we use reinforcement learning to create models of human behavior during assistance with dressing. To explore this kind of interaction, we assume that the robot presents an open sleeve of a hospital gown to a person, and that the person moves their arm into the sleeve. The controller that models the person's behavior is given the position of the end of the sleeve and information about contact between the person's hand and the fabric of the gown. We simulate this system with a human torso model that has realistic joint ranges, a simple robot gripper, and a physics-based cloth model for the gown. Through reinforcement learning (specifically the TRPO algorithm) the system creates a model of human behavior that is capable of placing the arm into the sleeve. We aim to model what humans are capable of doing, rather than what they typically do. We demonstrate successfully trained human behaviors for three robot-assisted dressing strategies: 1) the robot gripper holds the sleeve motionless, 2) the gripper moves the sleeve linearly towards the person from the front, and 3) the gripper moves the sleeve linearly from the side.

Alexander Clegg, Wenhao Yu, Zackory Erickson et al.

We present a controller that allows an arm-like manipulator to navigate deformable cloth garments in simulation through the use of haptic information. The main challenge of such a controller is to avoid getting tangled in, tearing or punching through the deforming cloth. Our controller aggregates force information from a number of haptic-sensing spheres all along the manipulator for guidance. Based on haptic forces, each individual sphere updates its target location, and the conflicts that arise between this set of desired positions is resolved by solving an inverse kinematic problem with constraints. Reinforcement learning is used to train the controller for a single haptic-sensing sphere, where a training run is terminated (and thus penalized) when large forces are detected due to contact between the sphere and a simplified model of the cloth. In simulation, we demonstrate successful navigation of a robotic arm through a variety of garments, including an isolated sleeve, a jacket, a shirt, and shorts. Our controller out-performs two baseline controllers: one without haptics and another that was trained based on large forces between the sphere and cloth, but without early termination.

Wenhao Yu, Jie Tan, C. Karen Liu et al.

We present a new method of learning control policies that successfully operate under unknown dynamic models. We create such policies by leveraging a large number of training examples that are generated using a physical simulator. Our system is made of two components: a Universal Policy (UP) and a function for Online System Identification (OSI). We describe our control policy as universal because it is trained over a wide array of dynamic models. These variations in the dynamic model may include differences in mass and inertia of the robots' components, variable friction coefficients, or unknown mass of an object to be manipulated. By training the Universal Policy with this variation, the control policy is prepared for a wider array of possible conditions when executed in an unknown environment. The second part of our system uses the recent state and action history of the system to predict the dynamics model parameters mu. The value of mu from the Online System Identification is then provided as input to the control policy (along with the system state). Together, UP-OSI is a robust control policy that can be used across a wide range of dynamic models, and that is also responsive to sudden changes in the environment. We have evaluated the performance of this system on a variety of tasks, including the problem of cart-pole swing-up, the double inverted pendulum, locomotion of a hopper, and block-throwing of a manipulator. UP-OSI is effective at these tasks across a wide range of dynamic models. Moreover, when tested with dynamic models outside of the training range, UP-OSI outperforms the Universal Policy alone, even when UP is given the actual value of the model dynamics. In addition to the benefits of creating more robust controllers, UP-OSI also holds out promise of narrowing the Reality Gap between simulated and real physical systems.