Speeding up 6-DoF Grasp Sampling with Quality-Diversity

TL;DR

This work tackles the data bottleneck in learning 6-DoF grasping by integrating Quality-Diversity (QD) optimization with robotic priors to rapidly generate diverse, robust grasps. It introduces a genotype-to-grasp projection that leverages synergies and priors, guided by a behavior-space-based QD archive to produce a large, diverse set of high-quality grasps. Across extensive simulations on four grippers and multiple YCB objects, QD-based sampling with priors outperforms standard priors and non-QD baselines, and transfers to real-world hardware with substantial sim-to-real fidelity. The results demonstrate the potential to build large, diverse grasp datasets that support robust and generalizable manipulation policies, with real-world experiments showing high transfer rates for the Panda and solid transfer for the Allegro. Overall, the approach offers a scalable path to data-efficient grasp policy learning by fast, diverse data generation that preserves sim-to-real transferability.

Abstract

Recent advances in AI have led to significant results in robotic learning, including natural language-conditioned planning and efficient optimization of controllers using generative models. However, the interaction data remains the bottleneck for generalization. Getting data for grasping is a critical challenge, as this skill is required to complete many manipulation tasks. Quality-Diversity (QD) algorithms optimize a set of solutions to get diverse, high-performing solutions to a given problem. This paper investigates how QD can be combined with priors to speed up the generation of diverse grasps poses in simulation compared to standard 6-DoF grasp sampling schemes. Experiments conducted on 4 grippers with 2-to-5 fingers on standard objects show that QD outperforms commonly used methods by a large margin. Further experiments show that QD optimization automatically finds some efficient priors that are usually hard coded. The deployment of generated grasps on a 2-finger gripper and an Allegro hand shows that the diversity produced maintains sim-to-real transferability. We believe these results to be a significant step toward the generation of large datasets that can lead to robust and generalizing robotic grasping policies.

Figures (8)

• Figure 1: Overview of the proposed framework. It consists of a population-based algorithm that allows the efficient exploration of the space of possible grasp positions from a genotype space. The optimization process is driven by searching for diverse and high-performing grasps using a structured archive of previously found solutions cully2022qd. The genotype space is designed to leverage the robotic priors that are commonly used in 6-DoF grasp sampling eppner2023abw2g. All grasps ever produced are added to an outcome archive, which is the output of the algorithm.
• Figure 2: Commonly used priors in 6DoF grasp sampling. (Left) approach-based ; (Right) antipodal-based. $R_g$ is the frame associated with the gripper. The cylinder is an illustration of a targeted object.
• Figure 3: Considered scenes. Grasp sampling schemes are evaluated on YCB objects calli2015benchmarking in the Pybullet coumans2016pybullet simulators. Experiments involve a FE Panda gripper, a Barrett hand, an Allegro hand and a Shadow hand.
• Figure 4: QD-based vs standard prior-based sampling schemes. Ratio of diverse, successful grasp found w.r.t. the number of evaluated samples. The dashed lines are commonly used in the literature; the plain ones are QD-based. contact_ME_scs outperforms the standard schemes and the raw QD method by a large margin on the 4 considered grippers.
• Figure 5: Leveraging priors with QD state-of-the-art. Same principle as in Fig. \ref{['fig:exp1_results']}. All the tested QD variants outperform standard prior-based sampling schemes. The pressure for quality optimization makes QD_contact variants behave similarly as QD_approach, such that the algorithm generates solutions that verify the usually hard-coded approach criterion (Fig. \ref{['fig:exp2_nu_results']}).