Dian Wang

Aiyuan Yang, Bin Xiao, Bingning Wang et al. · pku

Large language models (LLMs) have demonstrated remarkable performance on a variety of natural language tasks based on just a few examples of natural language instructions, reducing the need for extensive feature engineering. However, most powerful LLMs are closed-source or limited in their capability for languages other than English. In this technical report, we present Baichuan 2, a series of large-scale multilingual language models containing 7 billion and 13 billion parameters, trained from scratch, on 2.6 trillion tokens. Baichuan 2 matches or outperforms other open-source models of similar size on public benchmarks like MMLU, CMMLU, GSM8K, and HumanEval. Furthermore, Baichuan 2 excels in vertical domains such as medicine and law. We will release all pre-training model checkpoints to benefit the research community in better understanding the training dynamics of Baichuan 2.

Haojie Huang, Dian Wang, Xupeng Zhu et al.

Given point cloud input, the problem of 6-DoF grasp pose detection is to identify a set of hand poses in SE(3) from which an object can be successfully grasped. This important problem has many practical applications. Here we propose a novel method and neural network model that enables better grasp success rates relative to what is available in the literature. The method takes standard point cloud data as input and works well with single-view point clouds observed from arbitrary viewing directions.

Hai Nguyen, Andrea Baisero, Dian Wang et al.

Reinforcement learning in partially observable domains is challenging due to the lack of observable state information. Thankfully, learning offline in a simulator with such state information is often possible. In particular, we propose a method for partially observable reinforcement learning that uses a fully observable policy (which we call a state expert) during offline training to improve online performance. Based on Soft Actor-Critic (SAC), our agent balances performing actions similar to the state expert and getting high returns under partial observability. Our approach can leverage the fully-observable policy for exploration and parts of the domain that are fully observable while still being able to learn under partial observability. On six robotics domains, our method outperforms pure imitation, pure reinforcement learning, the sequential or parallel combination of both types, and a recent state-of-the-art method in the same setting. A successful policy transfer to a physical robot in a manipulation task from pixels shows our approach's practicality in learning interesting policies under partial observability.

Dian Wang, Jung Yeon Park, Neel Sortur et al.

Extensive work has demonstrated that equivariant neural networks can significantly improve sample efficiency and generalization by enforcing an inductive bias in the network architecture. These applications typically assume that the domain symmetry is fully described by explicit transformations of the model inputs and outputs. However, many real-life applications contain only latent or partial symmetries which cannot be easily described by simple transformations of the input. In these cases, it is necessary to learn symmetry in the environment instead of imposing it mathematically on the network architecture. We discover, surprisingly, that imposing equivariance constraints that do not exactly match the domain symmetry is very helpful in learning the true symmetry in the environment. We differentiate between extrinsic and incorrect symmetry constraints and show that while imposing incorrect symmetry can impede the model's performance, imposing extrinsic symmetry can actually improve performance. We demonstrate that an equivariant model can significantly outperform non-equivariant methods on domains with latent symmetries both in supervised learning and in reinforcement learning for robotic manipulation and control problems.

Haojie Huang, Dian Wang, Arsh Tangri et al.

Robotic pick and place tasks are symmetric under translations and rotations of both the object to be picked and the desired place pose. For example, if the pick object is rotated or translated, then the optimal pick action should also rotate or translate. The same is true for the place pose; if the desired place pose changes, then the place action should also transform accordingly. A recently proposed pick and place framework known as Transporter Net captures some of these symmetries, but not all. This paper analytically studies the symmetries present in planar robotic pick and place and proposes a method of incorporating equivariant neural models into Transporter Net in a way that captures all symmetries. The new model, which we call Equivariant Transporter Net, is equivariant to both pick and place symmetries and can immediately generalize pick and place knowledge to different pick and place poses. We evaluate the new model empirically and show that it is much more sample efficient than the non-symmetric version, resulting in a system that can imitate demonstrated pick and place behavior using very few human demonstrations on a variety of imitation learning tasks.

Xupeng Zhu, Dian Wang, Guanang Su et al.

Real-world grasp detection is challenging due to the stochasticity in grasp dynamics and the noise in hardware. Ideally, the system would adapt to the real world by training directly on physical systems. However, this is generally difficult due to the large amount of training data required by most grasp learning models. In this paper, we note that the planar grasp function is $\SE(2)$-equivariant and demonstrate that this structure can be used to constrain the neural network used during learning. This creates an inductive bias that can significantly improve the sample efficiency of grasp learning and enable end-to-end training from scratch on a physical robot with as few as $600$ grasp attempts. We call this method Symmetric Grasp learning (SymGrasp) and show that it can learn to grasp ``from scratch'' in less that 1.5 hours of physical robot time.

Dian Wang, Xupeng Zhu, Jung Yeon Park et al.

Although equivariant machine learning has proven effective at many tasks, success depends heavily on the assumption that the ground truth function is symmetric over the entire domain matching the symmetry in an equivariant neural network. A missing piece in the equivariant learning literature is the analysis of equivariant networks when symmetry exists only partially in the domain. In this work, we present a general theory for such a situation. We propose pointwise definitions of correct, incorrect, and extrinsic equivariance, which allow us to quantify continuously the degree of each type of equivariance a function displays. We then study the impact of various degrees of incorrect or extrinsic symmetry on model error. We prove error lower bounds for invariant or equivariant networks in classification or regression settings with partially incorrect symmetry. We also analyze the potentially harmful effects of extrinsic equivariance. Experiments validate these results in three different environments.

Hai Nguyen, Andrea Baisero, David Klee et al.

Incorporating inductive biases is a promising approach for tackling challenging robot learning domains with sample-efficient solutions. This paper identifies partially observable domains where symmetries can be a useful inductive bias for efficient learning. Specifically, by encoding the equivariance regarding specific group symmetries into the neural networks, our actor-critic reinforcement learning agents can reuse solutions in the past for related scenarios. Consequently, our equivariant agents outperform non-equivariant approaches significantly in terms of sample efficiency and final performance, demonstrated through experiments on a range of robotic tasks in simulation and real hardware.

Haojie Huang, Haotian Liu, Dian Wang et al.

Many manipulation tasks require the robot to rearrange objects relative to one another. Such tasks can be described as a sequence of relative poses between parts of a set of rigid bodies. In this work, we propose MATCH POLICY, a simple but novel pipeline for solving high-precision pick and place tasks. Instead of predicting actions directly, our method registers the pick and place targets to the stored demonstrations. This transfers action inference into a point cloud registration task and enables us to realize nontrivial manipulation policies without any training. MATCH POLICY is designed to solve high-precision tasks with a key-frame setting. By leveraging the geometric interaction and the symmetries of the task, it achieves extremely high sample efficiency and generalizability to unseen configurations. We demonstrate its state-of-the-art performance across various tasks on RLBench benchmark compared with several strong baselines and test it on a real robot with six tasks.

Xiaomeng Xu, Jisang Park, Han Zhang et al.

We present Whole-Body Mobile Manipulation Interface (HoMMI), a data collection and policy learning framework that learns whole-body mobile manipulation directly from robot-free human demonstrations. We augment UMI interfaces with egocentric sensing to capture the global context required for mobile manipulation, enabling portable, robot-free, and scalable data collection. However, naively incorporating egocentric sensing introduces a larger human-to-robot embodiment gap in both observation and action spaces, making policy transfer difficult. We explicitly bridge this gap with a cross-embodiment hand-eye policy design, including an embodiment agnostic visual representation; a relaxed head action representation; and a whole-body controller that realizes hand-eye trajectories through coordinated whole-body motion under robot-specific physical constraints. Together, these enable long-horizon mobile manipulation tasks requiring bimanual and whole-body coordination, navigation, and active perception. Results are best viewed on: https://hommi-robot.github.io

Dian Wang, Stephen Hart, David Surovik et al.

Recent work has shown diffusion models are an effective approach to learning the multimodal distributions arising from demonstration data in behavior cloning. However, a drawback of this approach is the need to learn a denoising function, which is significantly more complex than learning an explicit policy. In this work, we propose Equivariant Diffusion Policy, a novel diffusion policy learning method that leverages domain symmetries to obtain better sample efficiency and generalization in the denoising function. We theoretically analyze the $\mathrm{SO}(2)$ symmetry of full 6-DoF control and characterize when a diffusion model is $\mathrm{SO}(2)$-equivariant. We furthermore evaluate the method empirically on a set of 12 simulation tasks in MimicGen, and show that it obtains a success rate that is, on average, 21.9% higher than the baseline Diffusion Policy. We also evaluate the method on a real-world system to show that effective policies can be learned with relatively few training samples, whereas the baseline Diffusion Policy cannot.

Ondrej Biza, Dian Wang, Robert Platt et al.

In robotics, it is often not possible to learn useful policies using pure model-free reinforcement learning without significant reward shaping or curriculum learning. As a consequence, many researchers rely on expert demonstrations to guide learning. However, acquiring expert demonstrations can be expensive. This paper proposes an alternative approach where the solutions of previously solved tasks are used to produce an action prior that can facilitate exploration in future tasks. The action prior is a probability distribution over actions that summarizes the set of policies found solving previous tasks. Our results indicate that this approach can be used to solve robotic manipulation problems that would otherwise be infeasible without expert demonstrations. Source code is available at \url{https://github.com/ondrejba/action_priors}.

Yadong Li, Jun Liu, Tao Zhang et al.

We introduce Baichuan-Omni-1.5, an omni-modal model that not only has omni-modal understanding capabilities but also provides end-to-end audio generation capabilities. To achieve fluent and high-quality interaction across modalities without compromising the capabilities of any modality, we prioritized optimizing three key aspects. First, we establish a comprehensive data cleaning and synthesis pipeline for multimodal data, obtaining about 500B high-quality data (text, audio, and vision). Second, an audio-tokenizer (Baichuan-Audio-Tokenizer) has been designed to capture both semantic and acoustic information from audio, enabling seamless integration and enhanced compatibility with MLLM. Lastly, we designed a multi-stage training strategy that progressively integrates multimodal alignment and multitask fine-tuning, ensuring effective synergy across all modalities. Baichuan-Omni-1.5 leads contemporary models (including GPT4o-mini and MiniCPM-o 2.6) in terms of comprehensive omni-modal capabilities. Notably, it achieves results comparable to leading models such as Qwen2-VL-72B across various multimodal medical benchmarks.

Haojie Huang, Owen Howell, Dian Wang et al.

Many complex robotic manipulation tasks can be decomposed as a sequence of pick and place actions. Training a robotic agent to learn this sequence over many different starting conditions typically requires many iterations or demonstrations, especially in 3D environments. In this work, we propose Fourier Transporter (FourTran) which leverages the two-fold SE(d)xSE(d) symmetry in the pick-place problem to achieve much higher sample efficiency. FourTran is an open-loop behavior cloning method trained using expert demonstrations to predict pick-place actions on new environments. FourTran is constrained to incorporate symmetries of the pick and place actions independently. Our method utilizes a fiber space Fourier transformation that allows for memory-efficient construction. We test our proposed network on the RLbench benchmark and achieve state-of-the-art results across various tasks.

Xupeng Zhu, David Klee, Dian Wang et al.

Recent advances in Keyframe Imitation Learning (IL) have enabled learning-based agents to solve a diverse range of manipulation tasks. However, most approaches ignore the rich symmetries in the problem setting and, as a consequence, are sample-inefficient. This work identifies and utilizes the bi-equivariant symmetry within Keyframe IL to design a policy that generalizes to transformations of both the workspace and the objects grasped by the gripper. We make two main contributions: First, we analyze the bi-equivariance properties of the keyframe action scheme and propose a Keyframe Transporter derived from the Transporter Networks, which evaluates actions using cross-correlation between the features of the grasped object and the features of the scene. Second, we propose a computationally efficient coarse-to-fine SE(3) action evaluation scheme for reasoning the intertwined translation and rotation action. The resulting method outperforms strong Keyframe IL baselines by an average of >10% on a wide range of simulation tasks, and by an average of 55% in 4 physical experiments.

Haibo Zhao, Yu Qi, Boce Hu et al.

We present Generalizable Hierarchical Skill Learning (GSL), a novel framework for hierarchical policy learning that significantly improves policy generalization and sample efficiency in robot manipulation. One core idea of GSL is to use object-centric skills as an interface that bridges the high-level vision-language model and the low-level visual-motor policy. Specifically, GSL decomposes demonstrations into transferable and object-canonicalized skill primitives using foundation models, ensuring efficient low-level skill learning in the object frame. At test time, the skill-object pairs predicted by the high-level agent are fed to the low-level module, where the inferred canonical actions are mapped back to the world frame for execution. This structured yet flexible design leads to substantial improvements in sample efficiency and generalization of our method across unseen spatial arrangements, object appearances, and task compositions. In simulation, GSL trained with only 3 demonstrations per task outperforms baselines trained with 30 times more data by 15.5 percent on unseen tasks. In real-world experiments, GSL also surpasses the baseline trained with 10 times more data.

Arsh Tangri, Ondrej Biza, Dian Wang et al.

Sample efficiency is critical when applying learning-based methods to robotic manipulation due to the high cost of collecting expert demonstrations and the challenges of on-robot policy learning through online Reinforcement Learning (RL). Offline RL addresses this issue by enabling policy learning from an offline dataset collected using any behavioral policy, regardless of its quality. However, recent advancements in offline RL have predominantly focused on learning from large datasets. Given that many robotic manipulation tasks can be formulated as rotation-symmetric problems, we investigate the use of $SO(2)$-equivariant neural networks for offline RL with a limited number of demonstrations. Our experimental results show that equivariant versions of Conservative Q-Learning (CQL) and Implicit Q-Learning (IQL) outperform their non-equivariant counterparts. We provide empirical evidence demonstrating how equivariance improves offline learning algorithms in the low-data regime.

Haojie Huang, Karl Schmeckpeper, Dian Wang et al.

Humans can imagine goal states during planning and perform actions to match those goals. In this work, we propose Imagination Policy, a novel multi-task key-frame policy network for solving high-precision pick and place tasks. Instead of learning actions directly, Imagination Policy generates point clouds to imagine desired states which are then translated to actions using rigid action estimation. This transforms action inference into a local generative task. We leverage pick and place symmetries underlying the tasks in the generation process and achieve extremely high sample efficiency and generalizability to unseen configurations. Finally, we demonstrate state-of-the-art performance across various tasks on the RLbench benchmark compared with several strong baselines and validate our approach on a real robot.

Xupeng Zhu, Dian Wang, Ondrej Biza et al.

In planar grasp detection, the goal is to learn a function from an image of a scene onto a set of feasible grasp poses in $\mathrm{SE}(2)$. In this paper, we recognize that the optimal grasp function is $\mathrm{SE}(2)$-equivariant and can be modeled using an equivariant convolutional neural network. As a result, we are able to significantly improve the sample efficiency of grasp learning, obtaining a good approximation of the grasp function after only 600 grasp attempts. This is few enough that we can learn to grasp completely on a physical robot in about 1.5 hours.

Haojie Huang, Dian Wang, Robin Walters et al.

Transporter Net is a recently proposed framework for pick and place that is able to learn good manipulation policies from a very few expert demonstrations. A key reason why Transporter Net is so sample efficient is that the model incorporates rotational equivariance into the pick module, i.e. the model immediately generalizes learned pick knowledge to objects presented in different orientations. This paper proposes a novel version of Transporter Net that is equivariant to both pick and place orientation. As a result, our model immediately generalizes place knowledge to different place orientations in addition to generalizing pick knowledge as before. Ultimately, our new model is more sample efficient and achieves better pick and place success rates than the baseline Transporter Net model.

Dian Wang, Robin Walters, Xupeng Zhu et al.

Recently, a variety of new equivariant neural network model architectures have been proposed that generalize better over rotational and reflectional symmetries than standard models. These models are relevant to robotics because many robotics problems can be expressed in a rotationally symmetric way. This paper focuses on equivariance over a visual state space and a spatial action space -- the setting where the robot action space includes a subset of $\rm{SE}(2)$. In this situation, we know a priori that rotations and translations in the state image should result in the same rotations and translations in the spatial action dimensions of the optimal policy. Therefore, we can use equivariant model architectures to make $Q$ learning more sample efficient. This paper identifies when the optimal $Q$ function is equivariant and proposes $Q$ network architectures for this setting. We show experimentally that this approach outperforms standard methods in a set of challenging manipulation problems.

Dian Wang, Colin Kohler, Robert Platt

In the spatial action representation, the action space spans the space of target poses for robot motion commands, i.e. SE(2) or SE(3). This approach has been used to solve challenging robotic manipulation problems and shows promise. However, the method is often limited to a three dimensional action space and short horizon tasks. This paper proposes ASRSE3, a new method for handling higher dimensional spatial action spaces that transforms an original MDP with high dimensional action space into a new MDP with reduced action space and augmented state space. We also propose SDQfD, a variation of DQfD designed for large action spaces. ASRSE3 and SDQfD are evaluated in the context of a set of challenging block construction tasks. We show that both methods outperform standard baselines and can be used in practice on real robotics systems.

Dian Wang, Colin Kohler, Andreas ten Pas et al.

Assistive robot manipulators must be able to autonomously pick and place a wide range of novel objects to be truly useful. However, current assistive robots lack this capability. Additionally, assistive systems need to have an interface that is easy to learn, to use, and to understand. This paper takes a step forward in this direction. We present a robot system comprised of a robotic arm and a mobility scooter that provides both pick-and-drop and pick-and-place functionality for open world environments without modeling the objects or environment. The system uses a laser pointer to directly select an object in the world, with feedback to the user via projecting an interface into the world. Our evaluation over several experimental scenarios shows a significant improvement in both runtime and grasp success rate relative to a baseline from the literature [5], and furthermore demonstrates accurate pick and place capabilities for tabletop scenarios.