ReGIL: Retrieval-Guided Imitation Learning from a Single Demonstration

Authors*,
Institution Name
Institution
icon Data Code(TBD) arXiv

Rather than only using the demonstration once, ReGIL repeatedly queries it during training to support exploration, regularization data collection, and reward construction.

Abstract

Learning robot manipulation policies with deep neural networks from a single demonstration remains highly challenging, as even small deviations from the demonstrated trajectory can quickly compound into failure, while collecting substantial online interaction data is costly. We propose ReGIL, a retrieval-guided imitation learning framework that treats a single demonstration as an external memory. ReGIL repeatedly queries this static memory throughout training to simultaneously guide exploration, generate the regularization buffer, and construct rewards. Specifically, it computes rewards through local temporal alignment between the current trajectory and the retrieved segment, providing step-wise and informative feedback for policy improvement. We evaluate ReGIL on robotic manipulation tasks from the LIBERO and Meta-World benchmarks under the single demonstration setting. ReGIL outperforms prior baselines in both success rate and training efficiency. In real-robot experiments, using only one demonstration and less than one hour of online training, ReGIL achieves over 75% success rate across three manipulation tasks with randomness in both initial robot pose and target position. These results demonstrate that leveraging the single demonstration as reusable memory can provide more than static supervision for efficient robot learning.

Method

Research result visualization

ReGIL begins with a retrieval-and-replay exploration phase. A Vision Foundation Model (VFM) extracts visual features from both the agent’s current trajectories and the expert demonstration to retrieve task-relevant expert segments. These retrieved segments are used for action replay during the early exploration and similarity-based reward shaping throughout the whole process, providing dense reward signals for learning. During the online optimization phase, the policy is updated through environment interaction using the retrieval-based rewards together with a decaying imitation regularization term.

Experimental Result

We evaluate our approach in both simulation and real world under one demonstration setting to evaluate ReGIL.

Simulation

In simulation,we evaluate our approach on both Meta-World and LIBERO task suites. We report the mean and standard deviation of the success rate across 5 random seeds for online methods, while report the average success rate for offline methods for 50 trials.

Baseline comprision

Research result visualization

Ablation Study

Real Robot

We evaluate our method on three real-world tasks using a Franka Emika Panda manipulator: (1) Reach (a fundamental positioning task), (2) Insert (requiring high precision), and (3) Open(involving rich-contact dynamics)

Demonstration

Reach

Insert

Open

Fixed Target

Random Target

Failure Cases

Reach

Insert

Open

Additional Analysis

Sensitivity Study on Retrieval Parameters

Table: Effects of Temporal Stride(H),Candidate size(k), and Vision Encoder Selection.

Benchmark History (H) [k=5] Candidate number (k) [H=5] Encoder [k=5, H=5]
H=10 H=5 H=3 k=10 k=5 k=3 k=1 CLIP DINO R3M
MetaWorld 0.48 0.53 0.46 0.53 0.53 0.50 0.48 0.475 0.53 0.45
LIBERO 0.50 0.47 0.42 0.47 0.47 0.50 0.37 0.017 0.47 0.17
Mean 0.49 0.50 0.44 0.50 0.50 0.50 0.425 0.246 0.50 0.31

Computational Efficiency Analysis

Table: Latency Breakdown and Computational Efficiency.

Task Total Pipeline Cost (ms) Breakdown Cost (ms) & Ratio Max Control Freq. (Hz)
Pure DINO Encoding Pure Search + S-DTW
MetaWorld 31.61 ± 2.17 30.30 ± 1.84 (95.9%) 1.29 ± 0.43 (4.1%) 31.64
LIBERO 31.07 ± 1.04 29.61 ± 1.02 (95.3%) 1.44 ± 0.05 (4.6%) 32.18
Real Insert 32.85 ± 4.64 30.56 ± 3.22 (93.0%) 2.26 ± 2.38 (6.9%) 30.45
Real Reach 31.94 ± 3.92 30.09 ± 3.82 (94.2%) 1.83 ± 0.21 (5.7%) 31.31
Real Open 31.44 ± 1.32 29.32 ± 1.29 (93.3%) 2.09 ± 0.21 (6.7%) 31.81

Success Buffer Size Across Tasks

Table: Mean and standard deviation of samples collected in the success buffer. Maximum capacity is indicated in parentheses.

Environment Suite (Maximum) Task Name Success Buffer Size
(Mean ± Std)
Simulation MetaWorld
(5000)		 button_press 1716.00 ± 266.17
door_open 2841.50 ± 674.94
drawer_close 3222.80 ± 244.98
drawer_open 2301.00 ± 243.07
hammer 455.83 ± 146.58
window_close 3031.20 ± 544.19
window_open 1807.60 ± 500.73
Simulation LIBERO
(10000)		 kitchen1_bowl_on_cabinet 9897.83 ± 34.32
kitchen1_drawer 9915.80 ± 22.56
kitchen1_bowl_on_plate 3023.83 ± 973.09
kitchen2_bowl_on_plate 7505.67 ± 959.24
living5_mug_on_plate 6228.40 ± 765.61
study4 9961.17 ± 25.16
Real World (1000) reach 814.00
(1300) insert 961.00
(1300) open 947.00

Success Buffer Qualitative Analysis

Task Trajectories Visualization

K1_Bowl_on_Cabinet

K2_Bowl_on_Plate

K1_Bowl_on_Plate

Research result visualization

Training performance comparison across three similar kitchen manipulation tasks. Left: Success rates across 5 runs, and the shaded regions indicate the standard deviation. Right: Average success buffer size collected during training.

BibTeX

@article{YourPaperKey2024,
  title={ReGIL: Retrieval-Guided Imitation Learning from a Single Demonstration},
  author={First Author and Second Author and Third Author},
  journal={Conference/Journal Name},
  year={2026},
  url={https://regil2026.github.io}
}