How can robots learn dexterous grasping skills efficiently and apply them adaptively based on user instructions? This work tackles two key challenges: efficient skill acquisition from limited human demonstrations and context-driven skill selection. We introduce AdaDexGrasp, a framework that learns a library of grasping skills from a single human demonstration per skill and selects the most suitable one using a vision-language model (VLM). To improve sample efficiency, we propose a trajectory following reward that guides reinforcement learning (RL) toward states close to a human demonstration while allowing flexibility in exploration. To learn beyond the single demonstration, we employ curriculum learning, progressively increasing object pose variations to enhance robustness. At deployment, a VLM retrieves the appropriate skill based on user instructions, bridging low-level learned skills with high-level intent. We evaluate AdaDexGrasp in both simulation and real-world settings, showing that our approach significantly improves RL efficiency and enables learning human-like grasp strategies across varied object configurations. Finally, we demonstrate zero-shot transfer of our learned policies to a real-world PSYONIC Ability Hand, with a 90% success rate across objects, significantly outperforming the baseline.

Recent advancements in differentiable simulators highlight the potential of policy optimization using simulation gradients. Yet, these approaches are largely contingent on the continuity and smoothness of the simulation, which precludes the use of certain simulation engines, such as Mujoco. To tackle this challenge, we introduce the adaptive analytic gradient. This method views the Q function as a surrogate for future returns, consistent with the Bellman equation. By analyzing the variance of batched gradients, our method can autonomously opt for a more resilient Q function to compute the gradient when encountering rough simulation transitions. We also put forth the Adaptive-Gradient Policy Optimization (AGPO) algorithm, which leverages our proposed method for policy learning. On the theoretical side, we demonstrate AGPO’s convergence, emphasizing its stable performance under non-smooth dynamics due to low variance. On the empirical side, our results show that AGPO effectively mitigates the challenges posed by non-smoothness in policy learning through differentiable simulation.