Xinghao Zhu

Mingxiao Huo, Mingyu Ding, Chenfeng Xu et al. · berkeley

Humans inherently possess generalizable visual representations that empower them to efficiently explore and interact with the environments in manipulation tasks. We advocate that such a representation automatically arises from simultaneously learning about multiple simple perceptual skills that are critical for everyday scenarios (e.g., hand detection, state estimate, etc.) and is better suited for learning robot manipulation policies compared to current state-of-the-art visual representations purely based on self-supervised objectives. We formalize this idea through the lens of human-oriented multi-task fine-tuning on top of pre-trained visual encoders, where each task is a perceptual skill tied to human-environment interactions. We introduce Task Fusion Decoder as a plug-and-play embedding translator that utilizes the underlying relationships among these perceptual skills to guide the representation learning towards encoding meaningful structure for what's important for all perceptual skills, ultimately empowering learning of downstream robotic manipulation tasks. Extensive experiments across a range of robotic tasks and embodiments, in both simulations and real-world environments, show that our Task Fusion Decoder consistently improves the representation of three state-of-the-art visual encoders including R3M, MVP, and EgoVLP, for downstream manipulation policy-learning. Project page: https://sites.google.com/view/human-oriented-robot-learning

Xinghao Zhu, Siddarth Jain, Masayoshi Tomizuka et al.

Vision-based tactile sensors typically utilize a deformable elastomer and a camera mounted above to provide high-resolution image observations of contacts. Obtaining accurate volumetric meshes for the deformed elastomer can provide direct contact information and benefit robotic grasping and manipulation. This paper focuses on learning to synthesize the volumetric mesh of the elastomer based on the image imprints acquired from vision-based tactile sensors. Synthetic image-mesh pairs and real-world images are gathered from 3D finite element methods (FEM) and physical sensors, respectively. A graph neural network (GNN) is introduced to learn the image-to-mesh mappings with supervised learning. A self-supervised adaptation method and image augmentation techniques are proposed to transfer networks from simulation to reality, from primitive contacts to unseen contacts, and from one sensor to another. Using these learned and adapted networks, our proposed method can accurately reconstruct the deformation of the real-world tactile sensor elastomer in various domains, as indicated by the quantitative and qualitative results.

Xingyao Lin, Xinghao Zhu, Tianyi Lu et al.

The ultimate goal of embodied agents is to create collaborators that can interact with humans, not mere executors that passively follow instructions. This requires agents to communicate, coordinate, and adapt their actions based on human feedback. Recently, advances in VLAs have offered a path toward this goal. However, most current VLA-based embodied agents operate in a one-way mode: they receive an instruction and execute it without feedback. This approach fails in real-world scenarios where instructions are often ambiguous. In this paper, we address this problem with the Ask-to-Clarify framework. Our framework first resolves ambiguous instructions by asking questions in a multi-turn dialogue. Then it generates low-level actions end-to-end. Specifically, the Ask-to-Clarify framework consists of two components, one VLM for collaboration and one diffusion for action. We also introduce a connection module that generates conditions for the diffusion based on the output of the VLM. This module adjusts the observation by instructions to create reliable conditions. We train our framework with a two-stage knowledge-insulation strategy. First, we fine-tune the collaboration component using ambiguity-solving dialogue data to handle ambiguity. Then, we integrate the action component while freezing the collaboration one. This preserves the interaction abilities while fine-tuning the diffusion to generate actions. The training strategy guarantees our framework can first ask questions, then generate actions. During inference, a signal detector functions as a router that helps our framework switch between asking questions and taking actions. We evaluate the Ask-to-Clarify framework in 8 real-world tasks, where it outperforms existing state-of-the-art VLAs. The results suggest that our proposed framework, along with the training strategy, provides a path toward collaborative embodied agents.

Yuqi Xiang, Feitong Chen, Qinsi Wang et al.

The capability to transfer mastered skills to accomplish a range of similar yet novel tasks is crucial for intelligent robots. In this work, we introduce $\textit{Diff-Transfer}$, a novel framework leveraging differentiable physics simulation to efficiently transfer robotic skills. Specifically, $\textit{Diff-Transfer}$ discovers a feasible path within the task space that brings the source task to the target task. At each pair of adjacent points along this task path, which is two sub-tasks, $\textit{Diff-Transfer}$ adapts known actions from one sub-task to tackle the other sub-task successfully. The adaptation is guided by the gradient information from differentiable physics simulations. We propose a novel path-planning method to generate sub-tasks, leveraging $Q$-learning with a task-level state and reward. We implement our framework in simulation experiments and execute four challenging transfer tasks on robotic manipulation, demonstrating the efficacy of $\textit{Diff-Transfer}$ through comprehensive experiments. Supplementary and Videos are on the website https://sites.google.com/view/difftransfer

Lingfeng Sun, Devesh K. Jha, Chiori Hori et al. · berkeley

Designing robotic agents to perform open vocabulary tasks has been the long-standing goal in robotics and AI. Recently, Large Language Models (LLMs) have achieved impressive results in creating robotic agents for performing open vocabulary tasks. However, planning for these tasks in the presence of uncertainties is challenging as it requires \enquote{chain-of-thought} reasoning, aggregating information from the environment, updating state estimates, and generating actions based on the updated state estimates. In this paper, we present an interactive planning technique for partially observable tasks using LLMs. In the proposed method, an LLM is used to collect missing information from the environment using a robot and infer the state of the underlying problem from collected observations while guiding the robot to perform the required actions. We also use a fine-tuned Llama 2 model via self-instruct and compare its performance against a pre-trained LLM like GPT-4. Results are demonstrated on several tasks in simulation as well as real-world environments. A video describing our work along with some results could be found here.

Dingkun Guo, Yuqi Xiang, Shuqi Zhao et al.

Robotic grasping is a fundamental aspect of robot functionality, defining how robots interact with objects. Despite substantial progress, its generalizability to counter-intuitive or long-tailed scenarios, such as objects with uncommon materials or shapes, remains a challenge. In contrast, humans can easily apply their intuitive physics to grasp skillfully and change grasps efficiently, even for objects they have never seen before. This work delves into infusing such physical commonsense reasoning into robotic manipulation. We introduce PhyGrasp, a multimodal large model that leverages inputs from two modalities: natural language and 3D point clouds, seamlessly integrated through a bridge module. The language modality exhibits robust reasoning capabilities concerning the impacts of diverse physical properties on grasping, while the 3D modality comprehends object shapes and parts. With these two capabilities, PhyGrasp is able to accurately assess the physical properties of object parts and determine optimal grasping poses. Additionally, the model's language comprehension enables human instruction interpretation, generating grasping poses that align with human preferences. To train PhyGrasp, we construct a dataset PhyPartNet with 195K object instances with varying physical properties and human preferences, alongside their corresponding language descriptions. Extensive experiments conducted in the simulation and on the real robots demonstrate that PhyGrasp achieves state-of-the-art performance, particularly in long-tailed cases, e.g., about 10% improvement in success rate over GraspNet. Project page: https://sites.google.com/view/phygrasp

Yu Fang, Yue Yang, Xinghao Zhu et al.

Vision-language-action (VLA) models present a promising paradigm by training policies directly on real robot datasets like Open X-Embodiment. However, the high cost of real-world data collection hinders further data scaling, thereby restricting the generalizability of VLAs. In this paper, we introduce ReBot, a novel real-to-sim-to-real approach for scaling real robot datasets and adapting VLA models to target domains, which is the last-mile deployment challenge in robot manipulation. Specifically, ReBot replays real-world robot trajectories in simulation to diversify manipulated objects (real-to-sim), and integrates the simulated movements with inpainted real-world background to synthesize physically realistic and temporally consistent robot videos (sim-to-real). Our approach has several advantages: 1) it enjoys the benefit of real data to minimize the sim-to-real gap; 2) it leverages the scalability of simulation; and 3) it can generalize a pretrained VLA to a target domain with fully automated data pipelines. Extensive experiments in both simulation and real-world environments show that ReBot significantly enhances the performance and robustness of VLAs. For example, in SimplerEnv with the WidowX robot, ReBot improved the in-domain performance of Octo by 7.2% and OpenVLA by 21.8%, and out-of-domain generalization by 19.9% and 9.4%, respectively. For real-world evaluation with a Franka robot, ReBot increased the success rates of Octo by 17% and OpenVLA by 20%. More information can be found at: https://yuffish.github.io/rebot/

Xinghao Zhu, Devesh K. Jha, Diego Romeres et al. · berkeley

Automating the assembly of objects from their parts is a complex problem with innumerable applications in manufacturing, maintenance, and recycling. Unlike existing research, which is limited to target segmentation, pose regression, or using fixed target blueprints, our work presents a holistic multi-level framework for part assembly planning consisting of part assembly sequence inference, part motion planning, and robot contact optimization. We present the Part Assembly Sequence Transformer (PAST) -- a sequence-to-sequence neural network -- to infer assembly sequences recursively from a target blueprint. We then use a motion planner and optimization to generate part movements and contacts. To train PAST, we introduce D4PAS: a large-scale Dataset for Part Assembly Sequences (D4PAS) consisting of physically valid sequences for industrial objects. Experimental results show that our approach generalizes better than prior methods while needing significantly less computational time for inference.

Xuanlin Li, Tong Zhao, Xinghao Zhu et al.

Contact-rich bimanual manipulation involves precise coordination of two arms to change object states through strategically selected contacts and motions. Due to the inherent complexity of these tasks, acquiring sufficient demonstration data and training policies that generalize to unseen scenarios remain a largely unresolved challenge. Building on recent advances in planning through contacts, we introduce Generalizable Planning-Guided Diffusion Policy Learning (GLIDE), an approach that effectively learns to solve contact-rich bimanual manipulation tasks by leveraging model-based motion planners to generate demonstration data in high-fidelity physics simulation. Through efficient planning in randomized environments, our approach generates large-scale and high-quality synthetic motion trajectories for tasks involving diverse objects and transformations. We then train a task-conditioned diffusion policy via behavior cloning using these demonstrations. To tackle the sim-to-real gap, we propose a set of essential design options in feature extraction, task representation, action prediction, and data augmentation that enable learning robust prediction of smooth action sequences and generalization to unseen scenarios. Through experiments in both simulation and the real world, we demonstrate that our approach can enable a bimanual robotic system to effectively manipulate objects of diverse geometries, dimensions, and physical properties. Website: https://glide-manip.github.io/

John Z. Zhang, Maks Sorokin, Jan Brüdigam et al.

This paper presents a sim-to-real approach that enables legged robots to dynamically manipulate large and heavy objects with whole-body dexterity. Our key insight is that by performing test-time steering of a pre-trained whole-body control policy with a sample-based planner, we can enable these robots to solve a variety of dynamic loco-manipulation tasks. Interestingly, we find our method generalizes to a diverse set of objects and tasks with no additional tuning or training, and can be further enhanced by flexibly adjusting the cost function at test time. We demonstrate the capabilities of our approach through a variety of challenging loco-manipulation tasks on a Spot quadruped robot in the real world, including uprighting a tire heavier than the robot's nominal lifting capacity and dragging a crowd-control barrier larger and taller than the robot itself. Additionally, we show that the same approach can be generalized to humanoid loco-manipulation tasks, such as opening a door and pushing a table, in simulation. Project code and videos are available at \href{https://sumo.rai-inst.com/}{https://sumo.rai-inst.com/}.

Pengcheng Wang, Xinghao Zhu, Yuxin Chen et al.

Reinforcement Learning and Imitation Learning have achieved widespread success in many domains but remain constrained during real-world deployment. One of the main issues is the additional requirements that were not considered during training. To address this challenge, policy customization has been introduced, aiming to adapt a prior policy while preserving its inherent properties and meeting new task-specific requirements. A principled approach to policy customization is Residual Q-Learning (RQL), which formulates the problem as a Markov Decision Process (MDP) and derives a family of value-based learning algorithms. However, RQL has not yet been applied to policy gradient methods, which restricts its applicability, especially in tasks where policy gradient has already proven more effective. In this work, we first derive a concise form of Soft Policy Gradient as a preliminary. Building on this, we introduce Residual Policy Gradient (RPG), which extends RQL to policy gradient methods, allowing policy customization in gradient-based RL settings. With the view of RPG, we rethink the KL-regularized objective widely used in RL fine-tuning. We show that under certain assumptions, KL-regularized objective leads to a maximum-entropy policy that balances the inherent properties and task-specific requirements on a reward-level. Our experiments in MuJoCo demonstrate the effectiveness of Soft Policy Gradient and Residual Policy Gradient.

Kuan Fang, Yuxin Chen, Xinghao Zhu et al.

We present SAGA, a versatile and adaptive framework for visuomotor control that can generalize across various environments, task objectives, and user specifications. To efficiently learn such capability, our key idea is to disentangle high-level semantic intent from low-level visuomotor control by explicitly grounding task objectives in the observed environment. Using an affordance-based task representation, we express diverse and complex behaviors in a unified, structured form. By leveraging multimodal foundation models, SAGA grounds the proposed task representation to the robot's visual observation as 3D affordance heatmaps, highlighting task-relevant entities while abstracting away spurious appearance variations that would hinder generalization. These grounded affordances enable us to effectively train a conditional policy on multi-task demonstration data for whole-body control. In a unified framework, SAGA can solve tasks specified in different forms, including language instructions, selected points, and example demonstrations, enabling both zero-shot execution and few-shot adaptation. We instantiate SAGA on a quadrupedal manipulator and conduct extensive experiments across eleven real-world tasks. SAGA consistently outperforms end-to-end and modular baselines by substantial margins. Together, these results demonstrate that structured affordance grounding offers a scalable and effective pathway toward generalist mobile manipulation.

Xiang Zhang, Shiyu Jin, Changhao Wang et al.

This paper introduces a discrete-continuous action space to learn insertion primitives for robotic assembly tasks. Primitive is a sequence of elementary actions with certain exit conditions, such as "pushing down the peg until contact". Since the primitive is an abstraction of robot control commands and encodes human prior knowledge, it reduces the exploration difficulty and yields better learning efficiency. In this paper, we learn robot assembly skills via primitives. Specifically, we formulate insertion primitives as parameterized actions: hybrid actions consisting of discrete primitive types and continuous primitive parameters. Compared with the previous work using a set of discretized parameters for each primitive, the agent in our method can freely choose primitive parameters from a continuous space, which is more flexible and efficient. To learn these insertion primitives, we propose Twin-Smoothed Multi-pass Deep Q-Network (TS-MP-DQN), an advanced version of MP-DQN with twin Q-network to reduce the Q-value over-estimation. Extensive experiments are conducted in the simulation and real world for validation. From experiment results, our approach achieves higher success rates than three baselines: MP-DQN with parameterized actions, primitives with discrete parameters, and continuous velocity control. Furthermore, learned primitives are robust to sim-to-real transfer and can generalize to challenging assembly tasks such as tight round peg-hole and complex shaped electric connectors with promising success rates. Experiment videos are available at https://msc.berkeley.edu/research/insertion-primitives.html.

Xinghao Zhu, Yefan Zhou, Yongxiang Fan et al.

Robotic grasping for a diverse set of objects is essential in many robot manipulation tasks. One promising approach is to learn deep grasping models from large training datasets of object images and grasp labels. However, empirical grasping datasets are typically sparsely labeled (i.e., a small number of successful grasp labels in each image). The data sparsity issue can lead to insufficient supervision and false-negative labels and thus results in poor learning results. This paper proposes a Maximum Likelihood Grasp Sampling Loss (MLGSL) to tackle the data sparsity issue. The proposed method supposes that successful grasps are stochastically sampled from the predicted grasp distribution and maximizes the observing likelihood. MLGSL is utilized for training a fully convolutional network that generates thousands of grasps simultaneously. Training results suggest that models based on MLGSL can learn to grasp with datasets composing of 2 labels per image. Compared to previous works, which require training datasets of 16 labels per image, MLGSL is 8x more data-efficient. Meanwhile, physical robot experiments demonstrate an equivalent performance at a 90.7% grasp success rate on household objects.

Xinghao Zhu, Lingfeng Sun, Yongxiang Fan et al.

Proposing grasp poses for novel objects is an essential component for any robot manipulation task. Planning six degrees of freedom (DoF) grasps with a single camera, however, is challenging due to the complex object shape, incomplete object information, and sensor noise. In this paper, we present a 6-DoF contrastive grasp proposal network (CGPN) to infer 6-DoF grasps from a single-view depth image. First, an image encoder is used to extract the feature map from the input depth image, after which 3-DoF grasp regions are proposed from the feature map with a rotated region proposal network. Feature vectors that within the proposed grasp regions are then extracted and refined to 6-DoF grasps. The proposed model is trained offline with synthetic grasp data. To improve the robustness in reality and bridge the simulation-to-real gap, we further introduce a contrastive learning module and variant image processing techniques during the training. CGPN can locate collision-free grasps of an object using a single-view depth image within 0.5 seconds. Experiments on a physical robot further demonstrate the effectiveness of the algorithm.

Shiyu Jin, Xinghao Zhu, Changhao Wang et al.

Peg-in-hole assembly is a challenging contact-rich manipulation task. There is no general solution to identify the relative position and orientation between the peg and the hole. In this paper, we propose a novel method to classify the contact poses based on a sequence of contact measurements. When the peg contacts the hole with pose uncertainties, a tilt-then-rotate strategy is applied, and the contacts are measured as a group of patterns to encode the contact pose. A convolutional neural network (CNN) is trained to classify the contact poses according to the patterns. In the end, an admittance controller guides the peg towards the error direction and finishes the peg-in-hole assembly. Simulations and experiments are provided to show that the proposed method can be applied to the peg-in-hole assembly of different geometries. We also demonstrate the ability to alleviate the sim-to-real gap.

Yongxiang Fan, Xinghao Zhu, Masayoshi Tomizuka

Precision grasps with multi-fingered hands are important for precise placement and in-hand manipulation tasks. Searching precision grasps on the object represented by point cloud, is challenging due to the complex object shape, high-dimensionality, collision and undesired properties of the sensing and positioning. This paper proposes an optimization model to search for precision grasps with multi-fingered hands. The model takes noisy point cloud of the object as input and optimizes the grasp quality by iteratively searching for the palm pose and finger joints positions. The collision between the hand and the object is approximated and penalized by a series of least-squares. The collision approximation is able to handle the point cloud representation of the objects with complex shapes. The proposed optimization model is able to locate collision-free optimal precision grasps efficiently. The average computation time is 0.50 sec/grasp. The searching is robust to the incompleteness and noise of the point cloud. The effectiveness of the algorithm is demonstrated by experiments.