Real-World Reinforcement Learning with MPC Scaffolding for Dexterous Manipulation
Authors: Emek Barış Küçüktabak, Karankumar Patel, Zhaodong Yang, Jinda Cui, Kazuhiro Sasabuchi, Jun Takamatsu
Abstract
Real-world reinforcement learning (RL) offers a promising route to dexterous manipulation policies that can adapt directly from physical interaction, but learning is hindered by inefficient early exploration and costly failures. We propose a framework that uses sampling-based model predictive control (MPC) as scaffolding for real-world dexterous RL, providing structured prior experience and task-directed guidance during learning without human demonstrations or corrective actions. A small set of MPC trajectories is first used to populate an offline replay buffer and to pretrain the actor and critic. During online learning, MPC intermittently guides data collection while an off-policy Soft Actor-Critic learner trains from both prior MPC experience and newly collected physical interaction, with control gradually transitioning to the learned policy. On continuous in-hand rotation with a 16-DoF Allegro hand, initialized from 20 MPC trajectories collected in 12 minutes on hardware, the policy reaches 100% success after 7 minutes of online RL, with about three object drops on average during training. After 20 minutes of online learning, the policy achieves more than five times the rotation speed of the MPC controller and completes 1000 consecutive rotations without a drop. Ablations show complementary benefits from MPC-based pretraining, retained MPC experience, and online MPC guidance, while additional experiments demonstrate rapid adaptation to new object geometries and successful goal-conditioned reorientation.
Real-world learning for dexterous hands remains brittle because high-dimensional hand actions amplify imitation errors and make reinforcement-learning exploration prone to contact-breaking motion. While combining imitation learning (IL) with online reinforcement learning (RL) can reduce manual supervision, unconstrained exploration in raw hand-action spaces is sample-inefficient and risky for physical hardware. We introduce a latent motion prior module (\prior{}) that maps recent hand-action histories to a compact, history-conditioned latent prior and decodes continuous latent commands into executable high-dimensional hand targets. Built on this prior, \method{} is a three-stage real-world dexterous learning framework: it pretrains \prior{} from demonstrations, trains a visuomotor policy that predicts native arm commands and latent hand-action offsets, and improves the policy with online residual RL in the same latent hand-action space. This shared, decodable interface lets residual exploration make local corrections near demonstrated, contact-consistent hand motions rather than perturbing every finger joint independently. We evaluate \method{} on four real-robot dexterous manipulation tasks against raw, linear, and discrete hand-action interfaces. Starting from small task-specific demonstration sets, \method{} achieves a 56.25% average IL success rate and raises it to 98.75% after online RL, reaching 100% final success on three tasks and 95% on the remaining task.
A key challenge in contact-rich dexterous manipulation is the need to jointly reason over global geometry and nonsmooth contact dynamics. End-to-end policies bypass this complexity, but often require large amounts of data and transfer poorly from simulation to reality. We address the limitations with a simple insight: dexterous manipulation is inherently hierarchical--at a high level, a robot decides where to touch (geometry); at a low level it determines how to move the object through contact dynamics. Building on this insight, we propose a hierarchical RL--MPC framework in which a high-level reinforcement learning (RL) policy predicts a contact intention, a novel object-centric interface that specifies (i) an object-surface contact location and (ii) a post-contact object subgoal pose. Conditioned on the contact intention, a low-level contact-implicit model predictive control (MPC) optimizes local contact modes and real-time (re)plans through contact dynamics to generate robot actions that robustly move the object toward each subgoal. We evaluate the framework on non-prehensile tasks, including geometry-generalized pushing across diverse object shapes, pivoting/flipping-based object reorientation, and environment-assisted object repositioning. It achieves high success rate with substantially reduced data (10 times less than end-to-end baselines), highly robust performance, and zero-shot sim-to-real transfer.
We present a primitive-informed sampling-based model predictive control (MPC) framework for multi-fingered dexterous manipulation. Sampling-based MPC evaluates candidate control trajectories through forward simulation without requiring gradients through complex contact dynamics. However, directly sampling these trajectories in the high-dimensional joint space of a dexterous hand is inefficient and makes performance strongly dependent on the sampling distribution. Our framework biases sampling using low-dimensional manipulation primitives that encode coordinated finger motions, while simultaneously optimizing joint-level residuals to adapt these motions to the current hand-object configuration. Task-related rollout constraints reject infeasible trajectories during forward simulation, improving the effective use of the sampling budget. We evaluate the approach on a 16 DoF Allegro hand using a synchronized MuJoCo digital twin. Ablations show that both the primitive and residual are necessary for reliable continuous in-hand rotation, that increasing the sampling budget alone does not recover this coordination, and that rollout constraints substantially improve success rate. A primitive extracted for one object size transfers to other sizes and remains effective under model mismatch. The framework further supports grasping, object reorientation, and coordinated arm-hand manipulation, using primitives extracted from both a simulation-trained policy and human hand-motion data.
Emek Barış Küçüktabak, Karankumar Patel, Jinda Cui +3