Source-linked AI summary
Bridge Data: Boosting Generalization of Robotic Skills with Cross-Domain Datasets
Frederik Ebert, Yanlai Yang, Karl Schmeckpeper, Bernadette Bucher, Georgios Georgakis, Kostas Daniilidis, Chelsea Finn, Sergey Levine
TL;DR
Robotics lacks reusable, diverse datasets that support generalization across new tasks and domains. The paper introduces and evaluates a 7,200-demonstration bridge dataset spanning 71 tasks and 10 environments. Joint training improves performance on new-domain tasks, while limited target-domain data can import prior skills.
Problem
Robotics datasets are usually tied to specific platforms and domains, leaving practical reuse across new tasks and environments an open question.
Method
The paper collects a multi-task, multi-domain bridge dataset and evaluates it with imitation learning using joint training across bridge and target-domain data.
Results
50% success averaged over tasks versus around 22% for single-task policies, a 2× improvement, when jointly training on bridge and target data for entirely new tasks and domains.
Takeaways & Limitations
Shared multi-task and multi-domain datasets may help researchers improve target-domain generalization and import additional skills with modest in-domain data.
Takeaways & Limitations
The evaluation covers only a few distinct settings, including 5 environments at UC Berkeley.
Abstract
from arXiv · showhide
Robot learning holds the promise of learning policies that generalize broadly. However, such generalization requires sufficiently diverse datasets of the task of interest, which can be prohibitively expensive to collect. In other fields, such as computer vision, it is common to utilize shared, reusable datasets, such as ImageNet, to overcome this challenge, but this has proven difficult in robotics. In this paper, we ask: what would it take to enable practical data reuse in robotics for end-to-end skill learning? We hypothesize that the key is to use datasets with multiple tasks and multiple domains, such that a new user that wants to train their robot to perform a new task in a new domain can include this dataset in their training process and benefit from cross-task and cross-domain generalization. To evaluate this hypothesis, we collect a large multi-domain and multi-task dataset, with 7,200 demonstrations constituting 71 tasks across 10 environments, and empirically study how this data can improve the learning of new tasks in new environments. We find that jointly training with the proposed dataset and 50 demonstrations of a never-before-seen task in a new domain on average leads to a 2x improvement in success rate compared to using target domain data alone. We also find that data for only a few tasks in a new domain can bridge the domain gap and make it possible for a robot to perform a variety of prior tasks that were only seen in other domains. These results suggest that reusing diverse multi-task and multi-domain datasets, including our open-source dataset, may pave the way for broader robot generalization, eliminating the need to re-collect data for each new robot learning project.
I. INTRODUCTION
Robot learning commonly requires collecting data and retraining separately for each task and environment, limiting generalization and increasing cost. This paper investigates whether reusable multi-task, multi-domain bridge data can support transfer to new tasks and domains.
- Robots are typically trained from scratch for each new task and environment, increasing data-collection costs and limiting policy generalization.
- Reusable datasets have supported generalization in computer vision and NLP, but robotics datasets are usually tied to one platform and domain.
- The paper hypothesizes that effective robotic data reuse requires coverage of multiple tasks and multiple domains.
- The proposed bridge dataset contains 7,200 demonstrations of 71 tasks across 10 environments and is released for reuse.
- The study evaluates transfer for matching tasks, importing prior tasks into a target domain, and improving entirely new tasks in new domains.
II. RELATED WORK
Prior robot-learning datasets generally lack the combination of multiple tasks and multiple domains needed to study cross-task and cross-domain generalization. This work instead focuses on using generic multi-domain, multi-task datasets with existing algorithms to improve new-task generalization.
- Most prior deep visuomotor-learning work trains a single task in a single domain.
- The proposed dataset has the most tasks among compared datasets and is the only dataset with more than two tasks spanning many domains.
- Prior multi-task reinforcement-learning work showed that data from other tasks can improve generalization, but studied this within a single domain.
- Existing robot datasets such as MIME, DAML, and RoboTurk mostly feature a single domain, limiting their suitability for new-task generalization across domains.
- Human-collected multi-domain data has been used previously, but one cited study presents only two grasping tasks.
III. BRIDGE DATASETS
The paper introduces bridge datasets as large, diverse collections of robotic behaviors gathered across varied settings. It describes a specific teleoperated dataset of kitchen-themed manipulation tasks collected with a low-cost robotic arm.
- A bridge dataset is a large and diverse dataset of robotic behaviors collected across a range of settings.
- The described bridge dataset was collected by teleoperating a low-cost robotic arm.
- Its behaviors cover kitchen-themed manipulation tasks and varied viewpoints, lighting conditions, objects, and scenes.
A. Boosting Generalization via Bridge Datasets
The paper defines three bridge-data generalization scenarios and proposes that diverse settings, shared structure, and varied tasks support cross-domain skill transfer.
- A. Boosting Generalization via Bridge Datasets: The evaluation considers matching-behavior transfer, zero-shot transfer with target support, and boosting generalization of entirely new tasks.These scenarios progressively relax task and domain overlap requirements.
- A. Boosting Generalization via Bridge Datasets: In matching-behavior transfer, target-domain demonstrations for tasks also present in bridge data are combined with bridge data to improve performance and generalization.The target tasks must be present in the bridge dataset.
- A. Boosting Generalization via Bridge Datasets: In zero-shot transfer with target support, demonstrations for a few target-domain tasks connect bridge data to additional tasks not demonstrated in the target domain.The imported tasks must already appear in the bridge data.
- A. Boosting Generalization via Bridge Datasets: For new-task generalization, 50 demonstrations of a task absent from bridge data are combined with bridge data to improve performance and generalization.This scenario requires neither task nor domain matching.
- A. Boosting Generalization via Bridge Datasets: The proposed bridge datasets require varied settings, shared structure with target domains, and enough tasks to break unwanted task-domain correlations.The paper gives kitchen tasks versus construction robots as an example of insufficient structural similarity.
B. A Bridge Dataset of Large-Scale Kitchen Tasks
The authors build a reusable kitchen-task bridge dataset using an inexpensive, widely available robot and multi-camera VR demonstrations across many environments.
- B. A Bridge Dataset of Large-Scale Kitchen Tasks: Demonstrations are collected by applying tracked user motions to the robot end-effector through inverse kinematics.Images are captured concurrently from 3 to 5 cameras, including webcams and Intel RealSense depth cameras.
- B. A Bridge Dataset of Large-Scale Kitchen Tasks: The bridge dataset contains 7,200 demonstrations for 71 household kitchen tasks across 10 distinct toy-kitchen environments.Each task has between 50 and 300 demonstrations.
IV. USING BRIDGE DATA IN IMITATION LEARNING
The proof-of-concept uses task-conditioned behavioral cloning to combine bridge and target-domain data, with dataset rebalancing during joint training.
- IV. USING BRIDGE DATA IN IMITATION LEARNING: The experiments use imitation learning with task-conditioned behavioral cloning and a task-id input to distinguish tasks.Images are processed by a 34-layer ResNet before subsequent policy processing.
- IV. USING BRIDGE DATA IN IMITATION LEARNING: Pretraining is significantly less effective than joint training in the reported experiments.The authors note that bridge data can also be combined with target data in different ways.
- IV. USING BRIDGE DATA IN IMITATION LEARNING: For matching and zero-shot scenarios, training rebalances the roughly 10:1 bridge-to-target trajectory ratio to 70% bridge data and 30% target data.The passage introduces separate handling for the new-task scenario but does not provide its remaining details.
V. EXPERIMENTAL RESULTS
Experiments evaluate bridge-data training across three transfer scenarios and find gains for matching tasks, imported tasks, and entirely new tasks in target domains.
- V. EXPERIMENTAL RESULTS: The evaluation compares policies with and without bridge data across new target domains with limited target-domain data.The three scenarios correspond to matching behaviors, zero-shot transfer with target support, and new-task generalization.
- V. EXPERIMENTAL RESULTS: 66% success averaged over tasks was achieved with joint training, versus 14% for direct transfer, 28% for target-domain-only training, and 18% for single-task training.This result concerns tasks present in both bridge and target datasets.
- V. EXPERIMENTAL RESULTS: 44% success averaged over tasks with joint training exceeded 30% for direct transfer when importing tasks absent from the target-domain demonstrations.No target-domain-only baseline was available because those tasks had no target data.
- V. EXPERIMENTAL RESULTS: 50% success averaged over tasks was obtained for new tasks with joint training, compared with around 22% for single-task policies.Improvements occurred on 6 out of 10 tasks across three evaluation environments, yielding a 2× overall improvement.
- V. EXPERIMENTAL RESULTS: The new-task evaluations used target environments and tasks with structural similarity to the bridge setting.The passage identifies structural similarity between environments and tasks as important.
VI. CONCLUSION
The experiments show that diverse bridge data improves robotic skill generalization across multiple scenarios, including tasks seen in the bridge data and entirely new tasks. The dataset can also import prior skills into a target domain, although evaluation remains limited to five Berkeley environments.
- Bridge data improves performance across scenarios involving both bridge-seen tasks and entirely new tasks.
- Bridge data can serve as a generic tool for improving generalization in a user’s target domain.
- Bridge data can import prior tasks into a target domain, increasing the repertoire of available skills.
- The experiments evaluate only a few distinct settings: five environments at UC Berkeley.
- The released dataset is intended to help policies generalize to varied environments without repeatedly collecting large, exhaustive datasets.