8th Conference on Robot Learning (CoRL 2024)
Best Paper Award, NeurIPS 2023 TouchProcessing Workshop
Kelin Yu*, Yunhai Han*, Qixian Wang, Vaibhav Saxena, Danfei Xu, Ye Zhao (* Equal contribution)
Georgia Institute of Technology
21.movAbstract
Tactile sensing is critical to fine-grained, contact-rich manipulation tasks, such as insertion and assembly. An effective strategy to learn tactile-guided policy is to train neural networks that map tactile signal to control based on teleoperated demonstration data. However, to provide the teleoperated demonstration, human users often rely on visual feedback to control the robot. This creates a gap between the sensing modality used for controlling the robot (visual) and the modality of interest (tactile). To bridge this gap, we introduce MimicTouch, a novel framework for learning policies directly from demonstrations provided by human users with their hands. The key innovations are i) a human tactile data collection system which collects multi-modal tactile dataset for learning human's tactile-guided control strategy, ii) an imitation learning-based framework for learning human's tactile-guided control strategy through such data, and iii) an online residual RL framework to bridge the embodiment gap between the human hand and the robot gripper. Through comprehensive experiments, we highlight the efficacy of utilizing human's tactile-guided control strategy to resolve contact-rich manipulation tasks.
Illustration of our MimicTouch Framework. In part (a), we collect the multi-modal tactile feedback as human tactile demonstrations. In part (b), we learn compact low-dimensional tactile representations. In part (c), we derive a latent policy though a non-parametric imitation learning method. In part (d), we refine the offline policy through online residual reinforcement learning on a physical robot.
We evaluate the data-collecting efficiency and success rate for our insertion task. We find that the human hand is around 5.5 times faster than the teleoperation with spacemouse. Also, it is 2.2 times more accurate than the teleoperation system. The qualitative comparison is shown below.
teleopcollect.mp4Spacemouse Teleoperation System (19 traj/hr)
38.5% success rate for successful demo
hand_collection (2).movHand-guided Teleoperation System (44 traj/hr)
58.8% success rate for successful demo
hand1.mp4Human Tactile Demonstrations (104 traj/hr)
83.3% success rate for successful demo
We find that our framework can effectively learn from human tactile demonstrations and deploy the policy into the real robot. Below we show that we transfer the human policy into the real robot.
hand1.mp4Human Tactile Demonstration
initial.movPolicy Rollout
We qualitatively compare the policy learned from Human Tactile Demonstrations and Teleoperations, and find that the policy learned from Teleoperation data is easy to generate useless motions or motions without continuous contact in contact-rich manipulation tasks. Below is the qualitative result, and the quantitative training performance can be found in the paper.
teleopcollect.mp4Policy from Spacemouse Teleoperation
hand_rollout.movPolicy from Hand-guided Teleoperation
initial.movPolicy from Human Tactile Demonstrations
We conduct online real-world RL finetuning to bridge the embodiment gap between humans and robots and further enrich contact reasoning. We visualize the training process below.
training.movThree iterations of RL training
Shift 0.8 cm to four different directions, +-x and +-y
shift_left.mp4shift_right.mp4shift_above.movshift_blow.movTilting for 10 degrees or 20 degrees
10 degree.mov10 degree right.mov20_degree.mp420 degree right.movTwo-stage Dense Packing
Multi-Material Dense Packing (Soft)
two stage.mov2-stage close.MOVtissue.mp4soft close.movMulti-Material Dense Packing (Rigid)
Furniture Assembly (Insertion + Adjustment)
pen.movskrewing.mp4The threading skill is a manual script, which is used to test our policy whether can insert and adjust the object to the correct pose for threading.
We find that both tactile feedback and audio feedback should be helpful for the contact-rich manipulation task. If the system lacks tactile feedback, sometimes the robot cannot detect the contact and generates some dangerous behaviors; if the system lacks audio feedback, the robot cannot get information about the object's pose, so most of the time it cannot find the correct motion. The qualitative results have been shown below, and you can find more quantitative results in our paper.
tactile (2).mp4No Tactile Feedback, cannot find the correct action
audio.mp4No Audio Feedback, sometimes cannot detect contact
failure_limit.mp4Reach the joint limit because of large collisions
in tilting 20 degrees
failure_sponge.mp4The contact cannot be detected because the
sponge was too soft.
failure-tissue.mp4The contact cannot be detected due to the
shape change of tissue
failure_range.mp4Exceed the range because hole is too small
failure_thread.mp4Doesn't adjust to the appropriate pose to complete the threading
Bibtex
@inproceedings{
yu2024mimictouch,
title={MimicTouch: Leveraging Multi-modal Human Tactile Demonstrations for Contact-rich Manipulation},
author={Kelin Yu and Yunhai Han and Qixian Wang and Vaibhav Saxena and Danfei Xu and Ye Zhao},
booktitle={8th Annual Conference on Robot Learning},
year={2024},
url={https://openreview.net/forum?id=7yMZAUkXa4}
}