Physics-Based Dexterous Manipulations with Estimated Hand Poses and Residual Reinforcement Learning

Garcia-Hernando, Guillermo; Johns, Edward; Kim, Taekyun

doi:10.1109/iros45743.2020.9340947

Cited by 47 publications

(26 citation statements)

References 46 publications

Supporting

Mentioning

Contrasting

Unclassified

Order By: Relevance

“…To avoid big MoCap datasets, some methods use deep reinforcement learning (RL) for body-scene [5,45,46] or hand-object [6,11] interactions. These methods show promising results for navigating terrains with varying height and gaps [45,46], sitting on chairs [5,54], using a hammer and opening a door [11], and for in-hand object reorientation [6]. Generalization to new bodies, object geometry, and interaction types remains a challenge.…”

Section: Related Workmentioning

confidence: 99%

GOAL: Generating 4D Whole-Body Motion for Hand-Object Grasping

Taheri¹,

Choutas²,

Black³

et al. 2021

Preprint

View full text Add to dashboard Cite

Section: Related Workmentioning

confidence: 99%

GOAL: Generating 4D Whole-Body Motion for Hand-Object Grasping

Taheri¹,

Choutas²,

Black³

et al. 2021

Preprint

View full text Add to dashboard Cite

“…RL's applications in robotics suffer from expensive training data collection in real robots [8]. By offering cheaper and safer data collection environment, simulation has gained huge popularity for RL training [4], [9]- [11]. However, the issue of the reality gap still remains one of the main problems hindering the policy learned in simulation from transferring well to the real world.…”

Section: Related Workmentioning

confidence: 99%

DROID: Minimizing the Reality Gap using Single-Shot Human Demonstration

Tsai

Ding

et al. 2021

Preprint

Self Cite

View full text Add to dashboard Cite

Reinforcement learning (RL) has demonstrated great success in the past several years. However, most of the scenarios focus on simulated environments. One of the main challenges of transferring the policy learned in a simulated environment to real world, is the discrepancy between the dynamics of the two environments. In prior works, Domain Randomization (DR) has been used to address the reality gap for both robotic locomotion and manipulation tasks. In this paper, we propose Domain Randomization Optimization IDentification (DROID), a novel framework to exploit single-shot human demonstration for identifying the simulator's distribution of dynamics parameters, and apply it to training a policy on a door opening task. Our results show that the proposed framework can identify the difference in dynamics between the simulated and the real worlds, and thus improve policy transfer by optimizing the simulator's randomization ranges. We further illustrate that based on these same identified parameters, our method can generalize the learned policy to different but related tasks.

show abstract

“…Dexterous Hand Control Different approaches have been used for controlling dexterous hands. Learning-based methods most often resort to an anchored hand for in-hand manipulation tasks [6,30], which removes the complexity of generating collision-free trajectories, or rely on expert demonstrations [8,12,17,31,32], which can be costly to obtain. [32] collect expert trajectories via teleoperation, which they leverage in an RL setup to learn complex manipulation tasks.…”

Section: Related Workmentioning

confidence: 99%

“…[32] collect expert trajectories via teleoperation, which they leverage in an RL setup to learn complex manipulation tasks. [12] obtain noisy expert demonstrations from videos and use residual RL to correct the inputs for hand-object interaction tasks. This is achieved via a combination of taskspecific and imitation rewards.…”

Section: Related Workmentioning

confidence: 99%

“…Finally, synthesizing sequences of hand motion requires the generation of smooth and plausible trajectories. While prior work investigates the control of dexterous hands by learning from full demonstration trajectories [12,32], we address the generation of hand motion from only a single-frame grasp reference. This is a more challenging setting, because the generation of humanlike hand-object interaction trajectories without dense supervision is not straightforward.…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

D-Grasp: Physically Plausible Dynamic Grasp Synthesis for Hand-Object Interactions

Christen¹,

Kocabas²,

Aksan³

et al. 2021

Preprint

View full text Add to dashboard Cite

We introduce the dynamic grasp synthesis task: given an object with a known 6D pose and a grasp reference, our goal is to generate motions that move the object to a target 6D pose. This is challenging, because it requires reasoning about the complex articulation of the human hand and the intricate physical interaction with the object. We propose a novel method that frames this problem in the reinforcement learning framework and leverages a physics simulation, both to learn and to evaluate such dynamic interactions. A hierarchical approach decomposes the task into low-level grasping and high-level motion synthesis. It can be used to generate novel hand sequences that approach, grasp, and move an object to a desired location, while retaining human-likeness. We show that our approach leads to stable grasps and generates a wide range of motions. Furthermore, even imperfect labels can be corrected by our method to generate dynamic interaction sequences. Video is available at https://eth-ait.github.io/d-grasp/.

show abstract

Physics-Based Dexterous Manipulations with Estimated Hand Poses and Residual Reinforcement Learning

Cited by 47 publications

References 46 publications

GOAL: Generating 4D Whole-Body Motion for Hand-Object Grasping

GOAL: Generating 4D Whole-Body Motion for Hand-Object Grasping

DROID: Minimizing the Reality Gap using Single-Shot Human Demonstration

D-Grasp: Physically Plausible Dynamic Grasp Synthesis for Hand-Object Interactions

Contact Info

Product

Resources

About