DexDrummer: In-Hand, Contact-Rich, and Long-Horizon Dexterous Robot Drumming

Hung-Chieh Fang; Amber Xie; Jennifer Grannen; Kenneth Llontop; Dorsa Sadigh

cs.ROMar 23, 2026

DexDrummer: In-Hand, Contact-Rich, and Long-Horizon Dexterous Robot Drumming

Authors: Hung-Chieh Fang, Amber Xie, Jennifer Grannen, Kenneth Llontop, Dorsa Sadigh

Organizations: Stanford University

Abstract

Performing in-hand, contact-rich, and long-horizon dexterous manipulation remains an unsolved challenge in robotics. Prior hand dexterity works have considered each of these three challenges in isolation, yet do not combine these skills into a single, complex task. To further test the capabilities of dexterity, we propose drumming as a testbed for dexterous manipulation. Drumming naturally integrates all three challenges: it involves in-hand control for stabilizing and adjusting the drumstick with the fingers, contact-rich interaction through repeated striking of the drum surface, and long-horizon coordination when switching between drums and sustaining rhythmic play. We present DexDrummer, a hierarchical object-centric bimanual drumming policy trained in simulation with sim-to-real transfer. The framework reduces the exploration difficulty of pure reinforcement learning by combining trajectory planning with residual RL corrections for fast transitions between drums. A dexterous manipulation policy handles contact-rich dynamics, guided by rewards that explicitly model both finger-stick and stick-drum interactions. In simulation, we show our policy can play two styles of music: multi-drum, bimanual songs and challenging, technical exercises that require increased dexterity. Across simulated bimanual tasks, our dexterous, reactive policy outperforms a fixed grasp policy by 1.87x across easy songs and 1.22x across hard songs F1 scores. In real-world tasks, we show song performance across a multi-drum setup. DexDrummer is able to play our training song and its extended version with an F1 score of 1.0.

Figures & tables

Fig. 1: DexDrummer : A hierarchical dexterous drumming framework trained in simulation and deployed in the real world. Left: Training in simulation. Our policy decomposes the task into a hierarchical controller. A high-level policy generates parameterized motion primitives that produce drumstick trajectories from musical inputs. A low-level policy then uses residual RL to learn corrective arm and hand actions for accurate trajectory tracking during fast transitions. Middle: Contact-targeted rewards. To handle contact-rich dynamics, we design rewards that target two types of interactions. In-hand contacts encourage stable finger–stick manipulation through fingertip contact rewards, a fulcrum reward between the thumb and index finger, and an arm energy penalty that promotes finger-dominant control. External contacts address stick–drum interactions with a trajectory guidance reward to encourage drum strikes and a contact curriculum that gradually introduces drum contact during training. Right: Real-world drumming. Policies trained in simulation transfer zero-shot to the real robot, enabling dexterous multi-drum playing over long horizons.

Fig. 2: Drumming Environments. Our first simulation environment includes bimanual, multi-drum song-playing. Our second environment involves unimanual, uni-drum control for a high-speed technical exercise. Finally, in the real-world, we play songs with a drum pad and cymbal.

Reward	Formula	Explanation	Weight
In-Hand Contact
Fingertip	$\exp{(-1/(n_{\text{contacts}}+\varepsilon))}$	Make fingertips contact the stick	1.0
Fulcrum	$g\bigl((\|\|\mathbf{p}_{\text{thumb}}-\mathbf{p}_{\text{stick}}\|\|_{2}+\|\|\mathbf{p}_{\text{index}}-\mathbf{p}_{\text{stick}}\|\|_{2})/2\bigr)$	Position thumb and index finger to hold the fulcrum	1.0
Arm	$\lVert\tau_{\text{arm\_joints}}\rVert+\lVert v_{\text{arm\_joints}}\rVert$	Penalize arm movement to incentivize finger movement	0.03
External Contact
Trajectory	$\mathds{1}_{\text{is\_grasped}}\cdot g(\|\|\mathbf{p}_{\text{stick}}-\hat{\mathbf{p}}_{\text{stick}}\|\|_{2})$	Guide the stick with a reference trajectory	2.0

TABLE I: Reward functions and curriculum.

n_{\text{contacts}}

is the number of fingertip contacts,

\mathbf{p}_{\text{thumb}},\mathbf{p}_{\text{index}},\mathbf{p}_{\text{stick}}

are positions,

g(\cdot)

is a shaping function from [ 47 ] mapping distances to

[0,1]

, and

\mathds{1}

denotes indicator functions for grasp.

Observation	Dimension
Arm Proprioception	$7\times n_{\text{hand}}$
Hand Proprioception	$20\times n_{\text{hand}}$
Stick Head Proprioception	$3\times n_{\text{hand}}$
Stick Tail Proprioception	$3\times n_{\text{hand}}$
Trajectory Plan: Stick Head	$3\times n_{\text{hand}}\times L$
Trajectory Plan: Stick Tail	$3\times n_{\text{hand}}\times L$

TABLE II: Observation space.

L

denotes lookahead horizon.

Fig. 3: Bimanual Song-Playing Rollout. We visualize 6 frames across a single song trajectory, with lighter colored drums and cymbals corresponding to a hit. Every song requires multiple combinations of drums to be hit.

Fig. 4: Results for Dexterous Song-Playing. Left : Reactive grasp outperforms fixed grasp by a large margin in long-horizon contacts. Right : For more challenging songs requiring frequent drum-to-drum transitions, reactive grasp still improves performance, but with a smaller margin, primarily due to the reduced action space of fixed grasp.

Fig. 5: Results for Finger-Driven Control. Left: As tempo (beats per minute) increases, trajectory error for finger-driven control decreases and the gap to arm-driven control widens, showing the superior dexterity of finger-driven control. Energy consumption is also substantially lower compared to arm-driven control. Right: Without contact-targeted rewards, finger-driven control struggles to manage contact interactions effectively.

Fig. 6: Sample Rollouts for the Dexterous Exercise. For fast-playing exercises, the arm-driven policy moves to unnatural positions, leading to energy-intensive and potentially dangerous positions. Finger-driven control without the contact curriculum is unable to effectively learn dexterous control, whereas adding the contact curriculum leads to the most natural, effective drum hits. The lighter-colored drum head denotes that the drum is being hit by the stick.

Fig. 7: Results for Real-world Unimanual Drum Playing. We denote the cymbal as

c

and the drum pad as

d

. The short seen sequence used for training is

ccdd

. The long seen sequence is

ccccdddd

, while the long unseen sequence is

ccddccdd

. Open-loop replay achieves reasonable performance but cannot adapt to real-world dynamics. Our policy is able to play unseen sequences (e.g.,

d\rightarrow c

) by conditioning on the desired trajectory.

Fig. 8: Real World Rollouts. We visualize part of one closed-loop trajectory. The lighter colors for the cymbal and drum show when the robot hits them with the stick, and the arrows visualize the direction the stick was moving in. In this continuous rollout, we two cymbal hits and two drum hits. Notably, for the cymbal, which is not fixed and can rotate around the world vector, the robot strikes with a relatively loose grip (top). However, after hitting the drum pad, which is stable, the hand adjusts and forms a firmer grip (bottom), as shown by the movement of the index finger and thumb.

Explore similar work

Jul 15, 2025

Robot Drummer: Learning Rhythmic Skills for Humanoid Drumming

Humanoid robots have seen remarkable advances in dexterity, balance, and locomotion, yet their role in expressive domains such as music performance remains largely unexplored. Musical tasks, like drumming, present unique challenges such as split-second timing, rapid contacts, and multi-limb coordination over performances lasting minutes. In this paper, we introduce Robot Drummer, a simulation framework for humanoid drumming across a diverse repertoire of songs. We formulate humanoid drumming as the realization of timed contact events encoded as a Rhythmic Contact Chain. To handle the long-horizon nature of musical performance, we decompose each track into fixed-length segments and train a single policy across all segments in parallel using reinforcement learning. Through extensive experiments on over thirty popular tracks, our results demonstrate that Robot Drummer consistently achieves high F1 scores and enables efficient learning of long-horizon musical performances. The learned behaviors exhibit emergent human-like drumming strategies, such as cross-arm strikes, and adaptive stick assignments, demonstrating the potential of reinforcement learning to bring humanoid robots into the domain of creative musical performance. Project page: robotdrummer.github.io

Asad Ali Shahid, Francesco Braghin, Loris Roveda

IDSIA, SUPSI, Lugano, Switzerland · Department of Mechanical Engineering, Politecnico di Milano, Italy

Long-Horizon Robotic Manipulation Robot Skill Learning Robotics Simulation Humanoid Robot Control Robotic RL

Ask PDF

Mar 12, 2026

HandelBot: Real-World Piano Playing via Fast Adaptation of Dexterous Robot Policies

Mastering dexterous manipulation with multi-fingered hands has been a grand challenge in robotics for decades. Despite its potential, the difficulty of collecting high-quality data remains a primary bottleneck for high-precision tasks. While reinforcement learning and simulation-to-real-world transfer offer a promising alternative, the transferred policies often fail for tasks demanding millimeter-scale precision, such as bimanual piano playing. In this work, we introduce HandelBot, a framework that combines a simulation policy and rapid adaptation through a two-stage pipeline. Starting from a simulation-trained policy, we first apply a structured refinement stage to correct spatial alignments by adjusting lateral finger joints based on physical rollouts. Next, we use residual reinforcement learning to autonomously learn fine-grained corrective actions. Through extensive hardware experiments across five recognized songs, we demonstrate that HandelBot can successfully perform precise bimanual piano playing. Our system outperforms direct simulation deployment by a factor of 1.8x and requires only 30 minutes of physical interaction data.

Amber Xie, Haozhi Qi, Dorsa Sadigh

Stanford University · Amazon FAR (Frontier AI & Robotics)

Robotic Manipulation Bimanual Robotic Manipulation Sim-to-Real Transfer Data-Efficient Robot Learning

Ask PDF

Oct 3, 2026

PatternDex: Learning Interaction Patterns to Guide Reinforcement Learning of Bimanual Dexterous Manipulation of Articulated Objects

In this paper, we develop a method that enables bimanual dexterous hands to manipulate articulated objects with a high success rate without suffering from an embodiment gap. We observe that the correlation between hand motions and object motions is dictated by the object rather than the hands and can be learned from human-object demonstrations. Based on this observation, we propose PatternDex, a method that learns this correlation and represents it as a token sequence, which we call an interaction pattern. From this pattern, PatternDex estimates the wrist motions and contact points that fit the target robot, and then trains a reinforcement learning policy that exploits these estimates as guidance. Since the guidance fits the target embodiment, the policy explores only the actions that the target robot can execute and thus achieves high success rates. PatternDex also requires only simple fine-tuning to train a new robot, since it can reuse the learned interaction pattern. We evaluate PatternDex with bimanual dexterous hands on human demonstrations from the ARCTIC dataset. PatternDex achieves, on average, a 92.8% success rate with Allegro hands, while the state-of-the-art baseline achieves 52.2%. Also, it achieves success rates above 70% with three other robot hands after fine-tuning alone. Furthermore, we verify that the learned policy transfers well to a real-world task of opening a microwave. Videos and additional results are available at https://patterndex.github.io/PatternDex/

David Minkwan Kim, Runfa Blark Li, Beckham Po-Ju Lee +2

University of California, San Diego

Robotic Manipulation Bimanual Robotic Manipulation Articulated Object Manipulation Robotic RL Cross-Embodiment Robot Learning

Ask PDF