Haoqi Yuan 袁昊琦

Being-0: A Humanoid Robotic Agent with Vision-Language Models and Modular Skills

Haoqi Yuan, Yu Bai, Yuhui Fu, Bohan Zhou, Yicheng Feng, Xinrun Xu, Yi Zhan, Börje F. Karlsson, Zongqing Lu

arXiv / project page / GitHub / bibtex / blog

Being-0 is a hierarchical agent framework for humanoid robots, with a novel Vision-Language Model module bridging the gap between the Foundation Model's language-based task plans and the execution of low-level skills. Being-0 is capable of controlling humanoid robots with multi-fingered dexterous hands and active cameras, enhancing their dexterity in both navigation and manipulation tasks, and solving complex, long-horizon embodied tasks in the real world.

Cross-Embodiment Dexterous Grasping with Reinforcement Learning

Haoqi Yuan, Bohan Zhou, Yuhui Fu, Zongqing Lu

ICLR, 2025

conference paper / arXiv / project page / GitHub / bibtex / blog

CrossDex is an RL-based method for cross-embodiment dexterous grasping. Inspired by human teleoperation, we propose universal eigengrasp actions, which are converted to actions of various dexterous hands through retargeting. CrossDex successfully controls various hand embodiments with a single policy, and effectively transfers to unseen embodiments through zero-shot generalization and finetuning.

Efficient Residual Learning with Mixture-of-Experts for Universal Dexterous Grasping

Ziye Huang, Haoqi Yuan, Yuhui Fu, Zongqing Lu

ICLR, 2025

conference paper / arXiv / project page / GitHub / bibtex / blog

ResDex, our RL-based framework for universal dexterous grasping, achieves SOTA performance on DexGraspNet. ResDex employs residual policy learning for efficient multi-task RL, equipped with a mixture of geometry-unaware base policies that enhances generalization and diversity in grasping styles. ResDex masters grasping 3200 objects within 12-hours' training on a single RTX 4090 GPU, achieving zero generalization gaps.

Pre-Trained Multi-Goal Transformers with Prompt Optimization for Efficient Online Adaptation

Haoqi Yuan, Yuhui Fu, Feiyang Xie, Zongqing Lu

NeurIPS, 2024

conference paper / GitHub / bibtex

Previous works in skill pre-training utilize offline, task-agnostic dataset to accelerate RL. However, these approaches still require substantial RL steps to learn a new task. We propose MGPO, a method that leverages the power of Transformer-based policies to model sequences of goals during offline pre-training, enabling efficient online adaptation through prompt optimization.

RL-GPT: Integrating Reinforcement Learning and Code-as-policy

Shaoteng Liu, Haoqi Yuan, Minda Hu, Yanwei Li, Yukang Chen, Shu Liu, Zongqing Lu, Jiaya Jia

NeurIPS oral, 2024

conference paper / arXiv / project page / bibtex

RL-GPT equips Large Language Models (LLMs) with Reinforcement Learning (RL) tools, empowering LLM agents to solve challenging tasks in complex, open-world environments. It has a hierarchical framework: a slow LLM agent plans subtasks and selects proper tools (RL or code-as-policy); a fast LLM agent instantiates RL training pipelines or generates code to learn subtasks. LLM agents can perform self-improvement via trial-and-error efficiently. RL-GPT shows great efficiency on solving diverse Minecraft tasks, obtaining Diamond at 8% success rate within 3M environment steps.

Pre-Training Goal-Based Models for Sample-Efficient Reinforcement Learning

Haoqi Yuan, Zhancun Mu, Feiyang Xie, Zongqing Lu

ICLR oral (acceptance rate: 1.2%), 2024

conference paper / project page / GitHub / bibtex / talk

PTGM pre-trains on task-agnostic datasets to accelerate learning downstream tasks with RL. The pre-trained models provide: 1. a low-level, goal-conditioned policy that can perform diverse short-term behaviors; 2. a discrete high-level action space consisting of clustered goals in the dataset; 3. a goal prior model that guides and stablize downstream RL to train the high-level policy. PTGM can extend to the complicated domain Minecraft with large datasets, showing great sample efficiency, task performance, interpretability, and generalization of the acquired low-level skills.

Plan4MC: Skill Reinforcement Learning and Planning for Open-World Minecraft Tasks

Haoqi Yuan , Chi Zhang, Hongcheng Wang, Feiyang Xie, Penglin Cai, Hao Dong, Zongqing Lu

NeurIPS FMDM Workshop, 2023

workshop paper / arXiv / project page / GitHub / bibtex / blog

Plan4MC is a multi-task agent in the open world Minecraft, solving long-horizon tasks via planning over basic skills. It acquire three types of fine-grained basic skills through reinforcement learning without demonstrations. With a skill graph pre-generated by the Large Language Model, the skill search algorithm plans and interactively selects policies to solve complicated tasks. Plan4MC accomplishes 40 diverse tasks in Minecraft and unlocks Iron Pickaxe in the Tech Tree.

Robust Task Representations for Offline Meta-Reinforcement Learning via Contrastive Learning

Haoqi Yuan, Zongqing Lu

ICML, 2022

conference paper / GitHub / bibtex / talk

Offline meta-RL is a data-efficient RL paradigm that learns from offline data to adapt to new tasks. We propose a contrastive learning framework for robust task representations in context-based offline meta-RL. Our method improves the adaptation performance on unseen tasks, especially when the context is out-of-distribution.

DMotion: Robotic Visuomotor Control with Unsupervised Forward Model Learned from Videos

Haoqi Yuan, Ruihai Wu, Andrew Zhao, Haipeng Zhang, Zihan Ding, Hao Dong

IROS, 2021

conference paper / arXiv / project page / GitHub / bibtex / blog / talk

We study learning world models from action-free videos. Our unsupervised learning method leverages spatial transformers to disentangle the motion of controllable agent, learns a forward model conditioned on the explicit representation of actions. Using a few samples labelled with true actions, our method achieves superior performance on video prediction and model predictive control tasks.

Learning Diverse Bimanual Dexterous Manipulation Skills from Human Demonstrations

Bohan Zhou, Haoqi Yuan, Yuhui Fu, Zongqing Lu

arXiv / bibtex / blog

BiDexHD is a unified and scalable RL framework to learn bimanual manipulation skills, automatically constructing tasks from human trajectories and employing a teacher-student framework to obtain a vision-based policy tackling similar tasks. We demonstrate mastering 141 tasks from TACO dataset with a success rate of 74.59%.

Offline Model-Based Skill Stitching

Penglin Cai, Feiyang Xie, Haoqi Yuan, Zongqing Lu

manuscript / bibtex

Given pre-trained short-term skills, how can we stitch them to solve long-horizon tasks accurately? We propose a data-efficient framework for offline, model-based skill stitching, enabling effective transitions between skills.

Creative Agents: Empowering Agents with Imagination for Creative Tasks

Chi Zhang, Penglin Cai, Yuhui Fu, Haoqi Yuan , Zongqing Lu

arXiv / project page / GitHub / bibtex / blog

Creative tasks are challenging for open-ended agents, where the agent should give novel and diverse task solutions. We propose creative agents with the ability of imagination and introduce several variants in implementation. We benchmark creative tasks in the challenging open-world game Minecraft and propose novel evaluation metrics utilizing GPT-4V. Creative agents are the first AI agents accomplishing diverse building creation in Minecraft survival mode.

DLGAN: Disentangling Label-Specific Fine-Grained Features for Image Manipulation

Guanqi Zhan, Yihao Zhao, Bingchan Zhao, Haoqi Yuan , Baoquan Chen, Hao Dong

arXiv / bibtex

The first work to utilize discrete multi-labels to control which features to be disentangled, and enable interpolation between two domains without using continuous labels. An end-to-end method to support image manipulation conditioned on both images and labels, enabling both smooth and immediate changes simultaneously.