Closing the Reality Gap Zero-Shot Sim-to-Real Deployment for Dexterous Force-Based Grasping and Manipulation

Jul 16, 2026

Upload by Jiven

Human-like dexterous hands with multiple fingers enable human-level manipulation, yet direct transfer of control policies to real hardware is hindered by contact-rich physics and imperfect actuation. The work presents a practical sim-to-real reinforcement learning framework using dense tactile feedback and joint torque sensing to explicitly regulate physical interactions. It introduces a fast tactile simulation, a current-to-torque calibration removing the need for torque sensors, and actuator dynamics modeling with randomized non-ideal effects.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages23

Reading Time58 min

Introduction

Sim-to-Real in Locomotion and Prior Methods

Gaps in Dexterous In-Hand Manipulation

Motivation and Proposed Sim-to-Real Recipe

arXiv :2601 .02778v2 [ cs .RO] 9 Jan 2026 Closing the Reality Gap: Zero-Shot Sim-to-Real Deployment for Dexterous Force-Based Grasping and Manipulation ByteDance Seed Full Author List in Contributions Abstract Human-like dexterous hands with multiple fingers offer human-level manipulation capabilities, but training control policies that can directly deploy on real hardware remains difficult due to contactrich physics and imperfect actuation. We close this gap with a practical sim-to-real reinforcement learning (RL) framework that utilizes dense tactile feedback combined with joint torque sensing to explicitly regulate physical interactions. To enable effective sim-to-real transfer, we introduce (i) a computationally fast tactile simulation that computes distances between dense virtual tactile units and the object via parallel forward kinematics, providing high-rate, high-resolution touch signals needed by RL; (ii) a current-to-torque calibration that eliminates the need for torque sensors on dexterous hands by mapping motor current to joint torque; and (iii) actuator dynamics modeling to bridge the actuation gaps with randomization of non-ideal effects such as backlash, torque–speed saturation. Using an asymmetric actor–critic PPO pipeline trained entirely in simulation, our policies deploy directly to a five-finger hand. The resulting policies demonstrated two essential skills: (1) command-based, controllable grasp force tracking, and (2) reorientation of objects in the hand, both of which were robustly executed without fine-tuning on the robot. By combining tactile and torque in the observation space with effective sensing/actuation modeling, our system provides a practical solution to achieve reliable dexterous manipulation. To our knowledge, this is the first demonstration of controllable grasping on a multi-finger dexterous hand trained entirely in simulation and transferred zero-shot on real hardware. Date: 22 December, 2025 Project Page: [https://dexmanip-seed.github.io/dexmanip](https://dexmanip-seed.github.io/dexmanip) 1 Introduction Reinforcement learning with sim-to-real has achieved remarkable success in locomotion community, leading to widespread industrial use cases of quadruped and humanoid robots. Pioneering works such as ANYmal’s agile locomotion [22] and soccer-playing humanoids [12, 39] have demonstrated that a systematic design of domain randomization techniques, combined with RL techniques such as privileged learning of actuator dynamics and injection of noise in sensory observations, can bridge the reality gap for complex whole-body locomotion skills. Research in adaptive legged locomotion [43] further shows that while standard Multilayer Perceptron (MLP) architectures sufficed for learning individual narrow control policies, achieving complex multi-skill behaviors often necessitates the orchestration of a mixture of specialized neural network experts. Regarding sim-to-real techniques, through methods like domain randomization, system identification, and learning actuator compensation model [7, 16 , 22 , 25], these technologies reduce the gap between the robot model itself and its interaction with the environment, enabling robots to move across complex terrains [5, 15 , 18 , 23 , 49], perform mobile manipulation [7, 11 , 17 , 29 , 35], and mimic human motion [16 , 26] . Recent advancements like ASAP [16] leverage residual action modeling to align simulated joint torques with real-world actuator responses, while UAN [7] introduces unsupervised actuator networks to compensate for unmodeled nonlinear properties exhibiting in the real robots. These approaches share a common paradigm: (1) physics-based simulation augmentation to cover hardware uncertainties (e.g., friction coefficients, motor saturation), (2) learning-based actuator modeling to replace traditional model-based analytical methods which are overly simplified and inaccurate, and (3) RL learning techniques (asymmetric actor-

FAQs

Why is sim-to-real reinforcement learning difficult for dexterous in-hand manipulation?

Contact-rich physics with many contacts and material variations, plus tight sensorimotor coupling between tactile feedback, joint states, and high-dimensional perception, make current simulation and sensing/actuation models insufficient.

What sensing and modeling components are introduced to enable effective sim-to-real transfer?

The framework uses dense tactile simulation with parallel forward kinematics, a current-to-torque calibration to avoid torque sensors, and actuator dynamics modeling with randomized non-ideal effects.

What does the trained policy achieve on a real five-finger hand without fine-tuning?

The policies demonstrate controllable command-based grasp force tracking and robust object reorientation, executed zero-shot on real hardware.

Jul 16, 2026

Upload by Jiven