Few Shot Imitation Learning for Robotic Manipulation

This direction is split into other directions: sample inefficiency in imitation learning. Its last version stays readable and citable here.

Version 1 of 1154 papers155 conceptsfrozen 2026-10-04

The central challenge in modern robotics is not just performing a task, but learning it efficiently. While deep learning has revolutionized computer vision and natural language, its application to physical robots is often throttled by "data hunger." Standard imitation learning—where a robot learns by mimicking a human—typically requires hundreds or thousands of expert demonstrations to master even a simple task like picking up a block.

To overcome this, researchers are developing frameworks that allow robots to learn from a handful of examples, or even a single visual demonstration. This shift toward sample-efficient imitation learning relies on three core strategies: embedding geometric priors, decomposing tasks into reusable primitives, and leveraging "imagined" or unstructured data to fill the gaps.

The Geometry of Efficiency

Standard neural networks are "blind" to geometric transformations. If a robot learns to pick up a mug facing north, it usually has no inherent understanding of how to pick up the same mug facing south. It must see demonstrations at every possible orientation to generalize.

To solve this, researchers like Dian Wang and Haojie Huang have pioneered the use of equivariant neural networks. In the Equivariant Transporter Net (ETN), the system treats rotational and translational symmetry as a mathematical guarantee rather than something to be learned from scratch. By baking these symmetries into the architecture, the robot can generalize a single demonstration to any pose.

Similarly, Anthony Simeonov and colleagues introduced Neural Descriptor Fields (NDFs), which represent objects as continuous 3D-aware fields. Because these fields are $SE(3)$-equivariant, a robot can learn to hang a mug on a rack from just five demonstrations and immediately succeed even if the mug is placed in an entirely new, out-of-distribution orientation.

Decomposing the Problem

Not all parts of a robotic task are equally difficult to learn. Moving a hand through empty space to reach a table is relatively simple, while the precise, "contact-rich" moment of grasping a delicate object is complex.

Frameworks like Multi-Task Trajectory Transfer (MT3), developed by Kamil Dreczkowski and Pietro Vitiello, exploit this by decomposing tasks into "alignment" and "interaction" phases. By using a retrieval-based approach for the alignment and focusing learning only on the interaction, they enabled a robot to learn 1,000 distinct tasks from one demonstration each.

Edward Johns and Norman Di Palo further refined this in their work on Retrieval, Alignment, and Replay. Instead of end-to-end behavioral cloning, which maps pixels directly to motor commands, they use a single demonstration as a template. The robot retrieves a similar past experience, aligns its current view to the demonstration, and replays the movement, achieving order-of-magnitude better efficiency than traditional methods.

Learning to Learn In-Context

Perhaps the most ambitious direction is Meta-Imitation Learning (MIL). First introduced by Chelsea Finn and Sergey Levine, MIL allows a robot to "learn how to learn." By training on a wide variety of tasks, the robot develops a policy that can adapt to a brand-new task using just one visual demonstration and a single gradient update.

Building on this, Vitalis Vosylius developed Instant Policy, which achieves "In-Context Imitation Learning." Much like a Large Language Model can perform a task based on a prompt, Instant Policy allows a robot to execute a novel manipulation immediately after seeing a demonstration, without any weight updates or further training. It achieves this by modeling the problem as a graph-based diffusion process trained on vast amounts of simulated data.

Swapping Experts for Play and Imagination

When expert demonstrations are too expensive to collect, researchers look for cheaper alternatives. Jyothish Pari and Lerrel Pinto demonstrated that Playful Interactions—unstructured, non-expert data—can be used to learn visual representations that outperform models trained on ImageNet.

When even play data is scarce, robots can "dream." Leonardo Barcellona and Davide Allegro introduced DREMA, which uses Gaussian Splatting to create "learnable digital twins." By taking a single real-world demonstration and applying physically-verified "imagined" transformations in a simulator, the robot can practice a task thousands of times in its "mind" before ever attempting it again in reality.

By combining these geometric, structural, and generative insights, the field is moving toward a future where robots are no longer programmed or exhaustively trained, but simply shown what to do.

Papers in this version

Showing 60 of 154 papers.

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_few_shot_imitation_learning_for_robotic__dfffc404,
  title = {Few Shot Imitation Learning for Robotic Manipulation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/few-shot-imitation-learning-for-robotic-manipulation/rdr_dfffc4043a1247e28310a33b54573fe4},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_few_shot_imitation_learning_for_robotic__dfffc404_v1,
  title = {Few Shot Imitation Learning for Robotic Manipulation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/few-shot-imitation-learning-for-robotic-manipulation/rdr_dfffc4043a1247e28310a33b54573fe4/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}