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

Apply to capture

Glossary

What Is GraphIRL (Graph Inverse Reinforcement Learning from Diverse Videos)?

GraphIRL learns a robot reward from diverse third-person videos by turning each frame into an object graph and matching task progress over time. Third-party; Firsthand is not affiliated.

Last reviewed: November 2026

Short answer

GraphIRL (Graph Inverse Reinforcement Learning) is the self-supervised method in arXiv 2207.14299. It builds an object-centric graph from each video frame, learns an embedding that captures task progression through temporal matching, and turns distance to a goal into a reward for reinforcement learning. It is third-party research and Firsthand is not affiliated.

01

What is GraphIRL, and what problem does it address?

GraphIRL is named in Graph Inverse Reinforcement Learning from Diverse Videos (arXiv 2207.14299) by Sateesh Kumar, Jonathan Zamora, Nicklas Hansen, Rishabh Jangir and Xiaolong Wang of UC San Diego; the project page lists the Conference on Robot Learning (CoRL) 2022. The authors argue that inverse reinforcement learning from third-person videos removes the need for manual reward design, but that most prior work trains on a restricted domain of videos. Their insight is that a task can be described by entity interactions that form a graph, and that this abstraction removes irrelevant information such as textures, giving more robust reward functions.

02

Where does Firsthand fit next to GraphIRL?

GraphIRL is third-party research. Firsthand is not affiliated with the authors, and no named model, GraphIRL included, was trained on Firsthand data. Firsthand does not claim to use X-MAGICAL, an xArm robot, Soft Actor-Critic or any tool the paper names. The paper works with third-person video; 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. Object graphs, embeddings, reward functions and trained policies are not Firsthand deliverables.

03

How does GraphIRL turn videos into a reward?

Per the paper, the method takes a dataset of video demonstrations of a task and assumes no access to action information. An off-the-shelf detector extracts object bounding boxes from each frame, and each object becomes a node in a graph. A Spatial Interaction Encoder, which the authors build on Interaction Networks, models how objects relate to one another. The encoder is trained with a Temporal Cycle Consistency (TCC) loss, so the graph embeddings align across videos in time.

The reward is then the negative scaled Euclidean distance between the current observation and a representative goal frame in that embedding space. The authors state that any standard reinforcement learning algorithm can use it; they use Soft Actor-Critic (SAC) and no environment reward.

04

What videos did the GraphIRL authors use?

The paper evaluates on X-MAGICAL, a simulated cross-embodiment benchmark, in Standard and Diverse variants, where Diverse also randomizes visual appearance. For robot manipulation with an xArm, it uses real-world videos of humans doing Reach, Push and Peg in Box. The authors report collecting 256 Reach and 162 Peg in Box demonstrations and using 198 Push videos from Schmeckpeper et al. Objects are found with a hand-and-object detector the paper describes as trained on a large dataset collected from YouTube. The project page links the code and a dataset folder; licences are Not stated in the sources read.

05

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

The abstract reports significant gains in robustness to diverse video demonstrations over previous approaches, and better results than manual reward design on a real robot pushing task. The second table repeats the authors’ real-robot success rates, measured over 15 trials per method. Firsthand has not verified them. The authors also state a limitation: collecting demonstrations still requires human labor, and removing it entirely remains an open problem.

06

How does GraphIRL compare with related entries on this site?

The third table sets GraphIRL beside four site entries. The paper frames GraphIRL as an extension of XIRL to diverse videos, replacing image features with a graph abstraction while keeping temporal cycle-consistency.

07

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

Per the paper, it needs demonstration videos of one task, an object detector, a goal frame, and a simulator or robot with a reinforcement learning stack, all the buyer’s own. The paper uses third-person video and does not describe egocentric capture. 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.

08

GraphIRL at a glance, as stated in the paper, project page and official code

GraphIRL at a glance, as stated in the paper, project page and official code
ItemWhat the sources state
PaperGraph Inverse Reinforcement Learning from Diverse Videos (arXiv 2207.14299)
VenueConference on Robot Learning (CoRL) 2022, per the project page
InputDiverse third-person video demonstrations; no action information assumed
RepresentationObject bounding boxes from an off-the-shelf detector, built into a graph with one node per object
Learning signalSelf-supervised; Temporal Cycle Consistency loss on a Spatial Interaction Encoder
RewardNegative scaled distance to a goal frame in the learned embedding space
Policy learningSoft Actor-Critic; no environment reward used
EvaluationX-MAGICAL (simulated) and xArm manipulation tasks in simulation and on a real robot
LicencesNot stated in the sources read
Firsthand relationshipNone

10

Real-robot success rates reported by the authors (15 trials per method; Firsthand has not verified them)

Real-robot success rates reported by the authors (15 trials per method; Firsthand has not verified them)
TaskXIRLEnvironment rewardGraphIRL
Push0.270.470.60
Reach0.260.930.86
Peg in Box0.060.600.53

11

GraphIRL beside related entries on the site

GraphIRL beside related entries on the site
Site entryWhat the entry statesHow GraphIRL differs (sources read)
XIRLLearns an embedding with temporal cycle-consistency and uses distance to a goal state as the rewardApplies cycle-consistency to an object graph instead of image features, to handle diverse videos
DVDLearns a robot reward by comparing a robot video with a human demo videoLearns a goal-distance reward from many demonstration videos, with no robot video comparison
HOLDLearns a goal-image reward from unlabeled human videos via time regression or time-contrastive embeddingsLearns from detected objects and their interactions with temporal cycle-consistency
HuDORFine-tunes a robot hand policy online, using rewards from object points tracked in one human videoLearns its reward from many videos and trains in simulation before real-robot deployment

12

Frequently asked questions

What is GraphIRL?
Per arXiv 2207.14299, GraphIRL is a self-supervised method that learns a reward function from diverse third-person videos using an object-centric graph abstraction. It is third-party research.
Why use a graph instead of raw images?
The authors say a task can be described by entity interactions that form a graph, and that this abstraction removes irrelevant information such as textures, giving more robust reward functions.
Does GraphIRL need robot actions or an environment reward?
No. The paper states it does not assume access to action information in the demonstrations and does not use any environment reward, relying solely on the learned reward.
What data did GraphIRL use?
Simulated X-MAGICAL demonstrations and real-world videos of humans doing Reach, Push and Peg in Box, per the paper. Dataset licences are Not stated in the sources read.
Was GraphIRL trained on Firsthand data?
No. Firsthand is not affiliated with the authors, and no named model was trained on Firsthand data. Reward functions and embeddings 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.