Now recruiting capture teams worldwide. Get paid to record first-person video of real work.

Apply to capture

Glossary

What Is Domain-Adaptive Meta-Learning for One-Shot Imitation from Observing Humans?

DAML meta-learns a prior from human and robot demonstrations, then learns a robot task from one human video. Third-party; Firsthand is not affiliated.

Last reviewed: November 2026

Short answer

Domain-adaptive meta-learning (DAML) is the method in arXiv 1802.01557. A robot first meta-trains on human and robot demonstrations from many tasks. It then learns a new task from one human video, despite shifts in viewpoint, scene and embodiment. A learned temporal loss adapts the policy without human action labels. Third-party research; Firsthand is not affiliated.

01

What is domain-adaptive meta-learning, and what problem does it address?

DAML is named in One-Shot Imitation from Observing Humans via Domain-Adaptive Meta-Learning (arXiv 1802.01557) by Tianhe Yu, Chelsea Finn, Sudeep Dasari, Annie Xie, Tianhao Zhang, Pieter Abbeel and Sergey Levine, published at Robotics: Science and Systems 2018. The authors ask how a robot can learn a behavior from a single video of a human, even when there is domain shift in perspective, environment and embodiment between the robot and the human. They say prior approaches hand-specified how human and robot actions correspond and often relied on explicit human pose detection systems. DAML instead learns that correspondence from data. See also the site’s imitation learning and embodiment gap entries.

02

Where does Firsthand fit next to DAML?

DAML is third-party research. Firsthand is not affiliated with the authors, and no named model, DAML included, was trained on Firsthand data. Firsthand does not claim to use a PR2, a Sawyer arm, MuJoCo or any tool the paper names. The site states custom egocentric collection, fast and on demand: first-person human video with calibrated stereo depth, 3D hand pose at 21 joints per hand and 6-DoF IMUs at 200 Hz, with one task instruction per episode. Robot demonstrations, teleoperation data, meta-learning task sets and trained policies are not Firsthand deliverables.

03

How does domain-adaptive meta-learning learn from one human video?

Per the paper, DAML has two phases. In meta-training, the robot builds a prior over policies from human and robot demonstration data for a variety of tasks. At meta-test time it receives one human video of a new task and no robot demonstration, takes gradient steps from its learned initial parameters, and the resulting policy performs the task. The authors extend model-agnostic meta-learning (MAML) to handle the domain shift between a human demonstration and the robot, and say the key idea applies to other meta-learning algorithms.

04

What is the learned temporal adaptation objective?

A human video has no robot actions, and the human actions would not match the robot anyway, so a standard imitation loss cannot drive the update. The authors instead meta-learn an adaptation objective that uses only policy activations. During meta-training the policy is still supervised with true robot actions through a behavioral cloning objective. The learned objective is a small convolutional network with three layers of temporal convolutions, so it couples several frames of the video. The authors compare it with a linear per-timestep objective and state that the temporal version performs better.

05

What policy architecture does the paper use?

The policy is a convolutional network from RGB images to a distribution over actions. It has five convolutional layers, a spatial soft-argmax that extracts 2D feature points, and fully connected layers that also receive the robot configuration. The first convolutional layer is initialized from a network trained on ImageNet. Continuous actions use a mixture density network, and the gripper open or close action uses a sigmoid output. The paper does not use a hand or body pose detector; the authors say the approach is trained end to end.

06

What data did the authors meta-train on?

The authors collected their own demonstrations; no public human-video corpus is named. The first table lists what the paper states. The licence of this data is Not stated in the sources read. The human and robot demonstrations for a task must share the target object but need not be synchronized, start from matching object positions or have the same length.

07

What did the authors report, and what are the limits of this page?

The second table repeats figures the authors report; Firsthand has not verified them. The authors note that meta-training needs a large amount of demonstration data, a few thousand demonstrations from robots and humans in their experiments, though the amount per object is low, around 10 trials. They also say failures grew on novel backgrounds and that more diverse backgrounds or background augmentation would likely reduce them.

08

How does DAML differ from related methods on this site?

The third table sets DAML beside five site entries. The comparison uses only what those entries state and what the paper states.

09

What would a DAML-style approach need from custom collection?

Per the paper, the method needs human and robot demonstrations for many structurally similar tasks, with the robot demonstrations collected by the buyer on their own robot. The paper describes third-person observations of humans, not first-person video. Firsthand covers the human video side only, as custom collection at scale through a network of thousands of capture providers, for a consistent, recurring data flow to labs, confirmed on a Spec Pilot. The site’s 21-joint 3D hand pose is a signal the paper does not use.

10

DAML at a glance, as stated in the paper

DAML at a glance, as stated in the paper
ItemWhat the sources state
PaperOne-Shot Imitation from Observing Humans via Domain-Adaptive Meta-Learning (arXiv 1802.01557)
Method nameDAML (domain-adaptive meta-learning)
Authors and venueTianhe Yu, Chelsea Finn, Sudeep Dasari, Annie Xie, Tianhao Zhang, Pieter Abbeel, Sergey Levine; Robotics: Science and Systems 2018
Test-time inputOne video of a human performing the task; no robot demonstration and no human action labels
Built onModel-agnostic meta-learning (MAML) with a learned temporal adaptation objective
RobotsA 7-DoF PR2 arm with teleoperated demonstrations and a 7-DoF Sawyer arm with kinesthetic demonstrations
TasksPlacing, pushing and pick-and-place with new objects
Project pageThe paper points to a supplementary website; its address was not readable in the sources read, so it is not linked
LicencesNot stated in the sources read
Firsthand relationshipNone

12

Meta-training data stated by the authors (as written in the paper)

Meta-training data stated by the authors (as written in the paper)
ItemWhat the paper states
PR2 task dataA dataset with hundreds of objects: 1293 placing, 640 pushing and 1008 pick-and-place robot demonstrations, with an equal number of human demonstrations
Robot demonstrationsTeleoperation on the PR2; kinesthetic teaching on the Sawyer
Human demonstrationsRGB video of humans; from the robot camera perspective in the first experiment, and from a mounted cell-phone camera on ten table textures in the domain-shift experiment
Pairing ruleSame target object; not synchronized, no matching initial positions, no equal length needed
Public datasetsNone named

13

Figures reported by the authors (Firsthand has not verified them)

Figures reported by the authors (Firsthand has not verified them)
MeasureWhat the authors report
Held-out objects, PR215 placing, 12 pushing and 15 pick-and-place objects, three trials each, with a novel human demonstrator
Pushing under domain shiftOne-shot success of 81.8% on a seen background and 66.7% and 72.7% on two novel backgrounds
Sawyer placing77.8% success on 18 held-out objects, three trials each
Simulated pushing, no domain shift80.63% with the temporal objective versus 66.44% with the linear loss, in a MuJoCo setup

14

DAML beside related entries on the site

DAML beside related entries on the site
Site entryWhat the entry statesHow DAML differs (sources read)
Imitation learningTrains a policy to reproduce demonstrated behavior from example trajectoriesDAML meta-learns to infer a policy from one human video, using meta-training demonstrations from other tasks
AVIDTranslates human demonstration images into robot images with CycleGAN, then learns each stage with model-based reinforcement learningDAML does no image translation; the authors say it needs no additional robot experience or known model at test time
Embodiment gapThe mismatch between a human body and a robot’s hardwareDAML learns the human-to-robot correspondence implicitly instead of specifying it by hand
OKAMITeaches a humanoid a manipulation skill from a single RGB-D human videoDAML uses RGB video, a learned prior and PR2 and Sawyer arms
Time-Contrastive NetworksLearns a viewpoint-invariant embedding from unlabeled multi-view videoDAML’s learned objective is an adaptation loss over policy activations, not an embedding of frames

15

Frequently asked questions

What is domain-adaptive meta-learning (DAML)?
Per arXiv 1802.01557, it is a method that meta-learns a prior from human and robot demonstrations of many tasks, then learns a new robot task from one video of a human. It is third-party research.
Does DAML need human action labels?
No. The human video has no actions, so the authors meta-learn an adaptation objective that uses only policy activations. Robot actions supervise the policy during meta-training only.
How many human videos does the robot need for a new task?
One. The authors say the video need not come from the robot’s perspective or even from the same room.
What data did the authors use, and is it licensed?
Their own human and robot demonstrations on a PR2 and a Sawyer arm. The licence is Not stated in the sources read, and no public dataset is named.
Does DAML use first-person video?
The paper describes third-person observations of humans. Firsthand’s capture is first-person, so the two differ.
Was DAML trained on Firsthand data?
No. Firsthand is not affiliated with the authors, and no named model was trained on Firsthand data. Robot demonstrations, teleoperation data and meta-learning task sets are not Firsthand deliverables.
How can I request a fast, on-demand custom egocentric dataset for this research?
Request a custom dataset with your tasks and signals. Collection is fast and on demand, with a first validated batch in a median of 48 hours where coverage is deep.