Easy-IIL: Reducing Human Operational Burden in Interactive Imitation Learning via an Assistant Expert

arXiv:2603.12769 · cs.RO · Submitted 2026-03-13 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: Today's paper: "Easy-IIL: Reducing Human Operational Burden in Interactive Imitation Learning via an Assistant Expert".

Dev: The gist:

Rosa: First, who's behind it and why it matters.

Paper summary: Rosa: So we talked about the general idea of Easy-IIL, which is using an assistant expert to take over data collection instead of needing constant human involvement. What’s the actual thesis they are driving with this framework?

Dev: The authors propose Easy-IIL as a novel interactive imitation learning framework specifically designed to minimize that human operational effort by putting an assistant expert in place to handle the bulk of data collection while keeping IIL performance comparable.

Taro: So, it’s not just about monitoring passively, like some earlier work did; it’s about actively using an assistant expert to replace the human in most of the data gathering.

Rosa: And they claim that this is effective because structured model-based imitation approaches already show promise in low-data regimes with fewer demonstrations compared to end-to-end policies.

Dev: But they acknowledge that those end-to-end policies eventually get better as you feed them more data, which is why Easy-IIL seems positioned to keep up.

Taro: The paper outlines a specific data collection strategy involving an offline phase and an online phase, using a gating function to decide whether the human takes control or the assistant expert does the work.

Rosa: It breaks that down into two parts: first, offline data collection where the assistant expert generates demonstrations based on its own policy and the human monitors anomalies there.

Dev: Then there’s online data collection where a robot policy randomly switches between what it learned from a novice policy or from that assistant expert, with human intervention only when the task is nearly failing.

Taro: The offline dataset is defined as Dof f which includes data generated by both the human and the assistant expert, partitioned into done and rest.

Rosa: And then for online data collection, they define Don which mixes actions from the human, the assistant expert, and even a novice policy in some rounds.

Dev: The implementation details mention building this assistant expert on existing work for one-shot multi-stage robotic manipulation, and it uses techniques like Grounded-SAM to process the input.

Taro: The end-to-end chunk-based IL policy itself is implemented using a diffusion policy trained with Mean Squared Error as the loss function, which they formulate in a specific weighted training loss equation.

Rosa: So we’ve covered what Easy-IIL is trying to achieve: reducing human burden by using an assistant expert for most data collection while keeping performance comparable through careful offline and online data management.

Dev: The next part of this paper shows the results from extensive simulations and real-world experiments demonstrating that this framework significantly cuts down the human intervention rate by four to five times compared to standard IIL methods.

Taro: And user studies back up those simulation numbers, showing that the subjective workload on the human expert is also significantly reduced across both simulation tasks.

Conclusion: Rosa: So we’ve looked at how Easy-IIL uses an assistant expert to handle most of the data collection while only requiring a single demonstration from the human expert to start things off and then intervening in critical failure states.

Dev: The authors are basically saying that this approach successfully reduces the human operational burden by four to five times compared to standard IIL, while still maintaining performance comparable to other IIL baselines.

Taro: It’s about using those model-based imitation methods as a partner instead of needing the human on every single step of the data collection process.

Rosa: The paper is titled "Easy-IIL: Reducing Human Operational Burden in Interactive Imitation Learning via an Assistant Expert," and its authors are Zhang, Tang, Dong, Huang, Gu, and Zhang.

Dev: The implication for us is that if we want to scale up interactive learning systems where human interaction is costly in terms of time and effort, this framework offers a way to make the data collection phase much more manageable for people.

Taro: It suggests that we don't always have to rely on the human expert being present throughout every single interaction; sometimes an intelligent assistant can do most of the heavy lifting.

Rosa: The paper shows that by carefully preserving both offline and online data quality, Easy-IIL keeps the interactive imitation learning performance solid even when you’re delegating so much work to an AI expert.

Dev: That means we have a method here that aims to bridge the gap between high-quality, intensive IIL setups and practical applications where human time is limited.

Shenzhen Key Laboratory of Robotics and Computer Vision

cs.RO

Submitted: 2026-03-13

Updated: 2026-10-09

Project page: https://sites.google.com/view/easy-iil

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 86/100

The gist: The gist: Easy-IIL proposes an Interactive Imitation Learning (IIL) framework that significantly reduces human operational burden by utilizing an off-the-shelf model-based imitation methods as an

Key concepts

Assistant Expert: This is an off-the-shelf model-based imitation method used in Easy-IIL. It acts as a surrogate for the human expert, collecting and generating most of the necessary demonstrations autonomously during data collection phases.
Offline Data Collection: The initial phase where the human expert provides a single demonstration to initialize both the assistant expert and start gathering data. The assistant then generates more offline demonstrations while the human monitors for anomalies.
Online Data Collection: The stage where the robot policy interacts with an environment in multiple rounds. The system uses a gating function to decide whether to use the novice policy or the assistant expert, with human intervention only occurring when near task failure is imminent.

Terminology

Ablation — 9× in this paper · explained in 1 episode
Interactive Imitation Learning — 9× in this paper
Ablation Studies — 4× in this paper · explained in 1 episode
International Conference on Robotics and Automation — 3× in this paper
Ablation study — 2× in this paper · explained in 2 episodes
Computer Vision and Pattern Recognition — 2× in this paper
Foundation Model — 2× in this paper · explained in 3 episodes
Foundation Models — 2× in this paper · explained in 7 episodes
Frustration — 2× in this paper · explained in 2 episodes
Human-in-the-Loop — 2× in this paper · explained in 3 episodes
Reinforcement Learning — 2× in this paper · explained in 4 episodes
Aggregation — 1× in this paper · explained in 1 episode
Alignment — 1× in this paper · explained in 1 episode
Benchmark — 1× in this paper · explained in 2 episodes
Hyperparameters — 1× in this paper · explained in 2 episodes
Imitation Learning (IL) — 1× in this paper · explained in 2 episodes
Loss Function — 1× in this paper · explained in 2 episodes
Point Clouds — 1× in this paper · explained in 2 episodes
Regret — 1× in this paper · explained in 3 episodes
Reinforcement Learning (RL) — 1× in this paper · explained in 25 episodes
Robustness — 1× in this paper · explained in 5 episodes
Sample Efficiency — 1× in this paper · explained in 4 episodes
Sim-to-Real Gap — 1× in this paper · explained in 6 episodes
World Model — 1× in this paper · explained in 7 episodes

Summary

The gist: Easy-IIL proposes an Interactive Imitation Learning (IIL) framework that significantly reduces human operational burden by utilizing an off-the-shelf model-based imitation methods as an assistant expert to replace the human in most data collection while maintaining comparable IIL performance.

How it works

Easy-IIL is designed to minimize human operational effort by delegating the majority of data collection and interactions to an assistant expert, which substantially reduces active human operational effort<ref:2603.12769#pg6> The human expert only provides a single demonstration to initialize the assistant expert and intervenes in critical states where the task is approaching failure<ref:2603.12769#pg2> Furthermore, Easy-IIL can maintain IIL performance by preserving both offline and online data quality<ref:2603.12769#pg2>. Extensive simulation and real-world experiments demonstrate that EasyIIL significantly reduces human operational burden by 4–5 times compared to standard IIL, while maintaining comparable performance<ref:2603.12769#pg2>. User studies further confirm that the subjective workload on the human expert is significantly reduced<ref:2603.12769#pg2>.

Data Collection Strategy

The overall workflow begins with an offline phase where a human expert collects a single demonstration to initialize both an assistant expert and a dataset<ref:2603.12769#pg4>. Subsequently, the assistant expert collects the remaining demonstrations and adds them to the dataset, with human intervention occurring only when the task is on the verge of failure<ref:2603.12769#pg2>. The data collection strategy involves two main parts:

Offline Data Collection: In this phase, a human expert provides a single demonstration to initialize an assistant expert, which also serves as a dataset<ref:2603.12769#pg4>. Once initialized, the assistant expert is utilized to generate the remaining offline demonstrations while the human monitors this generation process, intervening only upon task anomalies<ref:2603.12769#pg4>. The offline dataset is defined as Dof f = (o, a)a ∈ (πH(x), πA(o)) and is partitioned into Done and Drest<ref:2603.12769#pg4>.
Online Data Collection: The online stage is divided into M rounds, where the robot policy in the i-th round πR i is defined by a gating function gi(j, k) that selects an action chunk from either the novice policy πN i-1 or the assistant expert πA randomly<ref:2603.12769#pg4>. Simultaneously, a human expert monitors the execution, providing intervention actions exclusively upon near task failure<ref:2603.12769#pg4>. The online dataset Don is defined as Don = (o, a)a ∈ (πH(x), πA(o), πN i-1(o)) and is composed of Dcorr1 ∪ Dcorr2 … ∪ DcorrM<ref:2603.12769#pg4>.

Implementation Details

The assistant expert is built upon existing works to achieve one-shot multi-stage robotic manipulation<ref:2603.12769#pg5>. For initialization, the human expert collects a single offline demonstration where each waypoint consists of an RGB-D image and the corresponding 6D end-effector (EE) pose and 1D gripper value<ref:2603.12769#pg5>. After initialization, the assistant expert takes the current RGB-D observation as input, employing Grounded-SAM and image matching techniques to process RGB-D images from both the demonstration and the current state<ref:2603.12769#pg5>.

The end-to-end chunk-based IL policy is implemented using a diffusion policy trained with Mean Squared Error (MSE) as the loss function<ref:2603.12769#pg5>. The weighted training loss is formulated as L:= Ea 0 k:k+Hp−1,ok,n,ϵ[Wk ⊗ (ϵ − ϵθ(a n k:k+Hp−1, ok, n)) 2], where the weight matrix Wk depends on the action’s source<ref:2603.12769#pg5>.

Experiment Results

Main experiments in RLBench demonstrate that Easy-IIL consistently achieves performance comparable to all IIL baselines, with success rate differences within a 3–4% margin<ref:2603.12769#pg7>. Crucially, Easy-IIL significantly reduces the human intervention rate across all tasks, achieving a substantial four- to five-fold reduction<ref:2603.12769#pg7>. Ablation results confirm the effectiveness of Easy-IIL in both stages; for instance, disabling novice actions in B significantly reduces human online intervention across both tasks<ref:2603.12769#pg6>.

User studies validate that Easy-IIL achieves a statistically significant reduction in subjective operational workload compared to IIL baselines<ref:2603.12769#pg8>. The Burden score, derived from the three metrics, is significantly lower for Easy-IIL compared to HG-DAgger and IWR across both simulation tasks<ref:2603.12769#pg8>.

Conclusion

Easy-IIL successfully reduces human operational burden while preserving comparable performance by leveraging an off-the-shelf assistant expert to replace the majority of human operations during data collection<ref:2603.12769#pg2>. The framework requires only a single offline demonstration from the human expert and intervention only when the task is on the verge of failure<ref:2603.12769#pg5>. Easy-IIL preserves both offline and online data quality to keep IIL performance comparable<ref:2603.12769#pg2>. Extensive simulation and real-world experiments demonstrate the effectiveness of Easy-IIL<ref:2603.12769#pg2>. Additionally, a user study confirms that Easy-IIL substantially lowers subjective operational workload<ref:2603.12769#pg8>. The proposed framework is an IIL framework for robotic manipulation that utilizes off-the-shelf model-based imitation methods as an assistant expert to replace a human expert in the majority of the data collection process in both offline and online stages<ref:2603.12769#pg2>. The human expert is only required to collect a single offline demonstration to initialize the assistant expert and perform intervention when the task is approaching failure<ref:2603.12769#pg5>. Easy-IIL preserves both offline and online data quality to keep IIL performance comparable<ref:2603.12769#pg2>. Extensive simulation and real-world experiments demonstrate that Easy-IIL reduces human operational burden by 4–5 times compared to standard IIL, while maintaining comparable performance<ref:2603.12769#pg2>. Furthermore, user studies validate that the subjective workload is significantly reduced compared to the baselines<ref:2603.12769#pg2>. The proposed framework is an IIL framework for robotic manipulation that utilizes off-the-shelf model-based imitation methods as an assistant expert to replace a human expert in the majority of the data collection process in both offline and online stages<ref:2603.12769#pg2>. The human expert is only required to collect a single offline demonstration to initialize the assistant expert and perform intervention when the task is approaching failure<ref:2603.12769#pg5>. Easy-IIL preserves both offline and online data quality to keep IIL performance comparable<ref:2603.12769#pg2>. Extensive simulation and real-world experiments demonstrate that Easy-IIL reduces human operational burden by 4–5 times compared to standard IIL, while maintaining comparable performance<ref:2603.12769#pg2>. Furthermore, user studies validate that the subjective workload is significantly reduced compared to the baselines<ref:2603.12769#pg2>. The proposed framework is an IIL framework for robotic manipulation that utilizes off-the-shelf model-based imitation methods as an assistant expert to replace a human expert in the majority of the data collection process in both offline and online stages<ref:2603.12769#pg2>. The human expert is only required to collect a single offline demonstration to initialize the assistant expert and perform intervention when the task is approaching failure<ref:2603.12769#pg5>. Easy-IIL preserves both offline and online data quality to keep IIL performance comparable<ref:2603.12769#pg2>.

Improvements for AI systems

textbfAssistant Expert Delegation for Data Collection Improvement: Leverage off-the-shelf model-based imitation methods as an assistant expert to replace active human operation for the majority of data collection in both offline and online stages (Easy-IIL, a framework that utilizes off-the-shelf model-based imitation methods as an assistant expert to replace active human operation for the majority of data collection). This allows the system to collect REST DEMO data autonomously while only requiring a human expert for ONE DEMO initialization and intervention in critical states.
textbfOnline Data Generation Strategy Refinement: Implement action chunk switching and noise injection to enhance exploration (Following [22], [26], [29], the manipulation workspace is divided into free space F and bottleneck space B, followed by we add zero-mean Gaussian noise with standard deviation σ to the normalized novice policy actions). This modification aims to enhance exploration for better online action quality.
textbfFailure State Mitigation via Workspace Prohibition: Disable novice policy execution in bottleneck space (we disable the execution of novice policy actions within B, which is implemented by Eq. (4) with xk ∈ B). Since B refers to a region near or in contact with the target object, this prevents low quality novice policy actions from occurring where precision is critical, thereby reducing failure rates and human intervention.
textbfrobust Policy Training via Weighted Loss Function: Utilize a weighted training loss formulation (the weighted training loss is formulated as: L:= Ea 0 k:k+Hp-1,ok,n,ϵ[Wk ⊗ (ϵ − ϵθ(a n k:k+Hp-1, ok, n))]). This ensures that we only utilize samples whose first action in the action chunk originates from either the assistant expert or the human expert, effectively biasing training toward high-quality demonstrations.
textbfPerformance Upper Bound Validation: Train an end-to-end policy to surpass the assistant expert (the end-toend IL policy trained via Easy-IIL achieves a higher performance upper bound than the model-based assistant expert). This confirms that while the assistant is efficient, the final learned policy can achieve superior performance at larger dataset scales.