Glossary
What Is XIRL (Cross-embodiment Inverse Reinforcement Learning)?
Last reviewed: November 2026
Short answer
XIRL (Cross-embodiment Inverse Reinforcement Learning) is the self-supervised method in arXiv 2106.03911. It uses temporal cycle-consistency constraints to learn visual embeddings that capture task progression from offline videos of different agents performing the same task. The negative distance between the current and goal state in that embedding space serves as a reward for reinforcement learning. Firsthand is not affiliated.
01
What is XIRL, and what problem does it address?
XIRL is named in XIRL: Cross-embodiment Inverse Reinforcement Learning (arXiv 2106.03911) by Kevin Zakka, Andy Zeng, Pete Florence, Jonathan Tompson, Jeannette Bohg and Debidatta Dwibedi; the project page lists Robotics at Google, Stanford University and UC Berkeley, and the paper appeared at the Conference on Robot Learning (CoRL) 2021. The authors study visual cross-embodiment imitation, where an agent learns from videos of other agents, such as humans, doing the same task despite differences in shape, actions and end-effector dynamics. They say that, before their work, producing rewards from self-supervised embeddings typically required alignment with a reference trajectory, which can be hard to acquire under stark embodiment differences. XIRL instead learns from videos of multiple expert agents, each doing the same task differently.
02
Where does Firsthand fit next to XIRL?
XIRL is third-party research. Firsthand is not affiliated with the authors, and no named model, XIRL included, was trained on Firsthand data. Firsthand does not claim to use X-MAGICAL, X-REAL, Soft Actor-Critic 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. Embeddings, reward functions, robot videos and trained policies are not Firsthand deliverables.
03
How does XIRL learn an embodiment-agnostic visual reward?
Per the abstract, XIRL is self-supervised. It applies temporal cycle-consistency constraints to offline demonstration videos from multiple expert agents and learns deep visual embeddings that capture task progression. The authors show that, with such task-progress-aware embeddings, simply taking the negative distance between the current state and the goal state in the embedding space is a useful reward for training policies with reinforcement learning. The official code names a Temporal Cycle Consistency trainer for pretraining and a Soft-Actor-Critic implementation adapted from pytorch_sac for the reinforcement learning stage.
04
What videos did the XIRL authors use?
The authors introduce X-MAGICAL, a simulated benchmark extension of MAGICAL for cross-embodiment imitation; the project page lists Gripper, Longstick, Mediumstick and Shortstick embodiments and a Sweeping task. They also collected X-REAL, which the project page describes as 93 demonstration videos of a human hand and 6 tools moving five pens into two cups. The repository README gives a download script for the X-MAGICAL demonstration dataset and says X-REAL will be released once approved; the sources read do not confirm its release. Dataset 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 that the learned reward works for embodiments seen during training and generalizes to entirely new embodiments, and that, when transferring real-world human demonstrations to a simulated robot, XIRL is more sample efficient than current best methods. Firsthand has not verified this. Only the abstract, project page and official README were read for this entry, so no scores or percentages are repeated.
06
How does XIRL compare with related entries on this site?
The second table sets XIRL beside four site entries. XIRL learns the reward itself, as embedding distance to a goal state, from offline videos of several embodiments.
07
What would an XIRL-style approach need from custom collection?
Per the sources read, it needs offline demonstration videos of several agents or end-effectors doing one task, a goal state, and a simulator or robot with a reinforcement learning stack, all the buyer’s own. The sources read do 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
XIRL at a glance, as stated in the paper, project page and official code
| Item | What the sources state |
|---|---|
| Paper | XIRL: Cross-embodiment Inverse Reinforcement Learning (arXiv 2106.03911) |
| Venue | Conference on Robot Learning (CoRL) 2021 |
| Input | Offline demonstration videos of multiple expert agents doing the same task |
| Learning signal | Self-supervised; temporal cycle-consistency constraints |
| Reward | Negative distance between the current and goal state in the learned embedding space |
| Policy learning | Reinforcement learning; the official code includes a Soft-Actor-Critic implementation adapted from pytorch_sac |
| Datasets | X-MAGICAL (simulated) and X-REAL (real world) |
| Licences | Not stated in the sources read |
| Firsthand relationship | None |
10
XIRL beside related entries on the site
| Site entry | What the entry states | How XIRL differs (sources read) |
|---|---|---|
| DVD | Learns a robot reward by comparing a robot video with a human demo video | Learns an embedding with temporal cycle-consistency and uses distance to a goal state as the reward |
| HOLD | Learns a goal-image reward from unlabeled human videos via time regression or time-contrastive embeddings | Learns from videos of multiple expert agents with temporal cycle-consistency |
| XSkill | Learns shared skill prototypes from unlabelled human and robot videos, then runs unseen tasks from one human prompt video | Learns a reward for reinforcement learning, not skill prototypes |
| HuDOR | Fine-tunes a robot hand policy online, using rewards from object points tracked in one human video | Learns its reward offline from many demonstration videos |
11
Frequently asked questions
- What is XIRL?
- Per arXiv 2106.03911, XIRL (Cross-embodiment Inverse Reinforcement Learning) is a self-supervised method that learns visual reward functions from videos of agents with different embodiments. It is third-party research.
- What does cross-embodiment mean in XIRL?
- The paper studies agents learning from videos of other agents, such as humans, doing the same task despite differences in shape, actions and end-effector dynamics.
- How does XIRL turn videos into a reward?
- It learns task-progress-aware embeddings with temporal cycle-consistency, then uses the negative distance between the current and goal state in that space as the reward.
- What data did XIRL use?
- The X-MAGICAL simulated benchmark and the real-world X-REAL videos, per the project page. Dataset licences are Not stated in the sources read.
- Was XIRL 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.
Related terms
- XSkill (Cross Embodiment Skill Discovery)
- Learning Reward Functions for Robotic Manipulation by Observing Humans
- DVD (Learning Generalizable Robotic Reward Functions from In-the-Wild Human Videos)
- HuDOR (Bridging the Human to Robot Dexterity Gap through Object-Oriented Rewards)
- Cross-embodiment learning
- Embodiment gap
- GraphIRL (Graph Inverse Reinforcement Learning from Diverse Videos)
- Time-Contrastive Networks (TCN)
Sources
- XIRL: Cross-embodiment Inverse Reinforcement Learning (arXiv 2106.03911)
- XIRL project page (official, linked from the paper)
- XIRL official code (google-research/xirl)
- Temporal Cycle-Consistency Learning (arXiv 1904.07846)
- Soft Actor-Critic: Off-Policy Maximum Entropy Deep Reinforcement Learning with a Stochastic Actor (arXiv 1801.01290)
- Learning Generalizable Robotic Reward Functions from "In-The-Wild" Human Videos (DVD, arXiv 2103.16817)
- XSkill: Cross Embodiment Skill Discovery (arXiv 2307.09955)
- Learning Reward Functions for Robotic Manipulation by Observing Humans (arXiv 2211.09019)