Bohan Wu

Junyu Luo, Bohan Wu, Xiao Luo et al. · pku

Post-training of Large Language Models (LLMs) is crucial for unlocking their task generalization potential and domain-specific capabilities. However, the current LLM post-training paradigm faces significant data challenges, including the high costs of manual annotation and diminishing marginal returns on data scales. Therefore, achieving data-efficient post-training has become a key research question. In this paper, we present the first systematic survey of data-efficient LLM post-training from a data-centric perspective. We propose a taxonomy of data-efficient LLM post-training methods, covering data selection, data quality enhancement, synthetic data generation, data distillation and compression, and self-evolving data ecosystems. We summarize representative approaches in each category and outline future research directions. By examining the challenges in data-efficient LLM post-training, we highlight open problems and propose potential research avenues. We hope our work inspires further exploration into maximizing the potential of data utilization in large-scale model training. Paper List: https://github.com/luo-junyu/Awesome-Data-Efficient-LLM

Bohan Wu, Suraj Nair, Li Fei-Fei et al.

In this paper, we study the problem of learning a repertoire of low-level skills from raw images that can be sequenced to complete long-horizon visuomotor tasks. Reinforcement learning (RL) is a promising approach for acquiring short-horizon skills autonomously. However, the focus of RL algorithms has largely been on the success of those individual skills, more so than learning and grounding a large repertoire of skills that can be sequenced to complete extended multi-stage tasks. The latter demands robustness and persistence, as errors in skills can compound over time, and may require the robot to have a number of primitive skills in its repertoire, rather than just one. To this end, we introduce EMBER, a model-based RL method for learning primitive skills that are suitable for completing long-horizon visuomotor tasks. EMBER learns and plans using a learned model, critic, and success classifier, where the success classifier serves both as a reward function for RL and as a grounding mechanism to continuously detect if the robot should retry a skill when unsuccessful or under perturbations. Further, the learned model is task-agnostic and trained using data from all skills, enabling the robot to efficiently learn a number of distinct primitives. These visuomotor primitive skills and their associated pre- and post-conditions can then be directly combined with off-the-shelf symbolic planners to complete long-horizon tasks. On a Franka Emika robot arm, we find that EMBER enables the robot to complete three long-horizon visuomotor tasks at 85% success rate, such as organizing an office desk, a file cabinet, and drawers, which require sequencing up to 12 skills, involve 14 unique learned primitives, and demand generalization to novel objects.

Bohan Wu, Feng Xu, Zhanpeng He et al.

Recent advances in deep reinforcement learning (RL) have demonstrated its potential to learn complex robotic manipulation tasks. However, RL still requires the robot to collect a large amount of real-world experience. To address this problem, recent works have proposed learning from expert demonstrations (LfD), particularly via inverse reinforcement learning (IRL), given its ability to achieve robust performance with only a small number of expert demonstrations. Nevertheless, deploying IRL on real robots is still challenging due to the large number of robot experiences it requires. This paper aims to address this scalability challenge with a robust, sample-efficient, and general meta-IRL algorithm, SQUIRL, that performs a new but related long-horizon task robustly given only a single video demonstration. First, this algorithm bootstraps the learning of a task encoder and a task-conditioned policy using behavioral cloning (BC). It then collects real-robot experiences and bypasses reward learning by directly recovering a Q-function from the combined robot and expert trajectories. Next, this algorithm uses the Q-function to re-evaluate all cumulative experiences collected by the robot to improve the policy quickly. In the end, the policy performs more robustly (90%+ success) than BC on new tasks while requiring no trial-and-errors at test time. Finally, our real-robot and simulated experiments demonstrate our algorithm's generality across different state spaces, action spaces, and vision-based manipulation tasks, e.g., pick-pour-place and pick-carry-drop.

Bohan Wu, Iretiayo Akinola, Jacob Varley et al.

Vision-based grasping systems typically adopt an open-loop execution of a planned grasp. This policy can fail due to many reasons, including ubiquitous calibration error. Recovery from a failed grasp is further complicated by visual occlusion, as the hand is usually occluding the vision sensor as it attempts another open-loop regrasp. This work presents MAT, a tactile closed-loop method capable of realizing grasps provided by a coarse initial positioning of the hand above an object. Our algorithm is a deep reinforcement learning (RL) policy optimized through the clipped surrogate objective within a maximum entropy RL framework to balance exploitation and exploration. The method utilizes tactile and proprioceptive information to act through both fine finger motions and larger regrasp movements to execute stable grasps. A novel curriculum of action motion magnitude makes learning more tractable and helps turn common failure cases into successes. Careful selection of features that exhibit small sim-to-real gaps enables this tactile grasping policy, trained purely in simulation, to transfer well to real world environments without the need for additional learning. Experimentally, this methodology improves over a vision-only grasp success rate substantially on a multi-fingered robot hand. When this methodology is used to realize grasps from coarse initial positions provided by a vision-only planner, the system is made dramatically more robust to calibration errors in the camera-robot transform.

Bohan Wu, Iretiayo Akinola, Peter K. Allen

Recent advances in on-policy reinforcement learning (RL) methods enabled learning agents in virtual environments to master complex tasks with high-dimensional and continuous observation and action spaces. However, leveraging this family of algorithms in multi-fingered robotic grasping remains a challenge due to large sim-to-real fidelity gaps and the high sample complexity of on-policy RL algorithms. This work aims to bridge these gaps by first reinforcement-learning a multi-fingered robotic grasping policy in simulation that operates in the pixel space of the input: a single depth image. Using a mapping from pixel space to Cartesian space according to the depth map, this method transfers to the real world with high fidelity and introduces a novel attention mechanism that substantially improves grasp success rate in cluttered environments. Finally, the direct-generative nature of this method allows learning of multi-fingered grasps that have flexible end-effector positions, orientations and rotations, as well as all degrees of freedom of the hand.