🤖 machine learning

GIFT: Geometry-Induced Functional Transfer for Category-level Object Manipulation

The paper presents GIFT, a framework that enables robots to generalize complex manipulation skills from a single human demonstration to new objects by leveraging functional maps for geometry-based interaction transfer and screw interpolation for smooth path generation.

Original authors: Cristiana de Farias, Luis Figueredo, Riddhiman Laha, Maxime Adjigble, Brahim Tamadazte, Rustam Stolkin, Sami Haddadin, Naresh Marturi

Published 2026-06-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Cristiana de Farias, Luis Figueredo, Riddhiman Laha, Maxime Adjigble, Brahim Tamadazte, Rustam Stolkin, Sami Haddadin, Naresh Marturi

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are teaching a robot how to shake a specific bottle to mix its contents. You grab the robot's arm and physically guide it through the motion: grabbing the bottle, shaking it, and setting it down.

Now, imagine you swap that bottle for a completely different one—a tall, skinny wine bottle instead of a short, fat soda bottle. A standard robot might get confused. It might try to grab the wine bottle in the exact same spot as the soda bottle (which would be the middle of the neck, causing a drop) or shake it in a way that doesn't fit the new shape.

This is the problem the paper GIFT (Geometry-Induced Functional Transfer) tries to solve. Here is how it works, explained simply:

1. The Core Idea: "The Map, Not the Territory"

Most robots learn by memorizing specific movements (like "move arm 5 inches left"). GIFT teaches the robot to understand the relationship between the robot's hand and the object's shape.

Think of it like this: Instead of teaching the robot "hold the bottle at point A," GIFT teaches it "hold the bottle where the surface feels like a handle." It separates the action (shaking) from the specific object (the soda bottle).

2. The Three Magic Ingredients

The paper uses three main tricks to make this happen:

A. The "Functional Map" (The Shape Translator)

Imagine you have a rubber sheet with a drawing of a smiley face on it. If you stretch that sheet into a different shape, the smiley face stretches with it, but the relationship between the eyes and the mouth stays the same.

GIFT uses a mathematical tool called Functional Maps to do exactly this with 3D objects.

The Demo: The robot sees the "handle" of the first bottle.
The Transfer: When it sees a new, differently shaped bottle, it uses this "rubber sheet" math to find the equivalent spot on the new bottle. It's like saying, "This new bottle has a 'handle' area just like the old one, even though the bottle looks totally different."

B. The "Screw" Motion (The Smooth Path)

Once the robot knows where to grab the new object, it needs to know how to move.

The Problem: If you just draw a line from point A to point B, the robot might bump into things or move awkwardly if the new object is in a different spot.
The Solution: GIFT uses something called ScLERP (Screw Linear Interpolation). Think of a screw. When you turn a screw, it moves forward and rotates at the same time in a perfect, smooth spiral.
The Analogy: Instead of just moving the robot's hand in a straight line, GIFT calculates the "screw path." This ensures that no matter how the new object is sitting on the table, the robot's hand moves in a smooth, natural curve that respects the physics of the original demonstration. It's like the robot remembers the "feel" of the motion, not just the coordinates.

C. The "One-Shot" Learning

Usually, if you want a robot to learn a new task, you have to show it hundreds of examples or spend hours training it. GIFT is a one-shot method.

The Claim: The robot only needs one single demonstration from a human. After seeing the action once on one object, it can immediately figure out how to do that same action on a completely different object of the same type (e.g., from one bottle to another bottle, or one box to another box) without any extra training.

3. How It Works in Real Life (The Experiment)

The researchers tested this with a 7-armed robot. They showed it how to:

Wipe a surface with a wooden block.
Draw a box with a marker.
Shake a bottle.
Press a button.

Then, they put the robot in front of different objects (different bottles, different blocks) that it had never seen before.

The Result: The robot successfully figured out where to grab the new objects and performed the tasks (wiping, shaking, pressing) correctly. It didn't need to be retrained. It just used the "map" to translate the old task to the new shape.

Summary

GIFT is like giving a robot a "universal translator" for physical tasks.

You show it once: "Do this with this object."
It analyzes the shape: It creates a mathematical map of where the interaction happens.
It adapts: When a new object appears, it finds the matching spot on the new shape using that map.
It moves smoothly: It uses "screw" math to move its arm naturally, ensuring it doesn't crash or drop the object, even if the new object is in a weird position.

The paper claims this makes robots much better at handling new, unfamiliar objects in messy, real-world environments without needing massive amounts of data or retraining.

Technical Summary: GIFT (Geometry-Induced Functional Transfer)

Problem Definition
Robotic manipulation in human-centered environments faces significant challenges when generalizing skills to unfamiliar objects or new scenes. While recent advances in vision-language-action (VLA) models and generative visuomotor policies (e.g., RT-2, Octo, Diffusion Policies) have improved zero-shot capabilities, they often lack explicit physical and geometric priors. Consequently, these methods can exhibit brittle behavior when tasks require strict constraint satisfaction or when the scene distribution shifts (e.g., changes in object shape, topology, or reference frames). Existing approaches often rely on large datasets, extensive retraining, or manual reference frames, failing to guarantee that contact, pose, or kinematic constraints are upheld during execution. The core problem addressed by this work is how to transfer complex manipulation skills from a single human demonstration to new, category-level objects with significantly different geometries while preserving task-defining constraints without additional training.

Methodology
The authors propose GIFT (Geometry-Induced Functional Transfer), a framework that treats geometric representation as the backbone of skill transfer rather than an auxiliary cue. The method operates in two stages: Demonstration and Imitation.

Skill Representation: A skill is defined as a tuple $T = \{X, f(\text{task})\}$ , where $X$ is the robot end-effector trajectory and $f(\text{task})$ is an interaction function defined over the object's surface. These functions encode:
- Robot Interaction Functions (RIF): Contact regions between the gripper and the object.
- Environment Interaction Functions (EIF): Interactions between the object and environmental elements (e.g., planes).
Functional Map Correspondence (FMC): To transfer skills to new objects, GIFT utilizes the Functional Maps framework. This involves:
- Computing orthonormal bases using the first $n$ eigenfunctions of the Laplace-Beltrami (LB) operator for both the demonstrated and target objects.
- Calculating descriptor functions (e.g., Wave Kernel Signature) that are invariant to isometries.
- Solving an optimization problem to find a functional map $C$ that aligns the descriptor functions of the source and target shapes.
- Refining this map into a point-to-point correspondence to transfer the interaction functions ( $f \to g$ ) to the new object's manifold.
- An object selection metric based on the diagonal dominance of the functional map matrix is used to identify the best-matching object in multi-object scenes.
Geometry-Consistent Trajectory Generation: Once the interaction functions are transferred, the robot generates a new trajectory using Screw Linear Interpolation (ScLERP).
- The demonstrated path is represented as a sequence of unit dual quaternions.
- Relative transformations ( $\delta_i$ ) between consecutive poses are computed.
- In the new scene, these relative transformations are applied to the new goal pose to reconstruct the path.
- This approach ensures ad-invariance, meaning the geometric constraints and relative relationships between points are preserved regardless of the reference frame or object topology, unlike decoupled interpolation methods.
Execution: The system selects the best-matching object, transfers the RIF and EIF, generates grasp poses using a planner (LoCoMo) constrained by the transferred RIF, and executes the task using the ScLERP-generated trajectory.

Key Contributions

One-Shot Skill Transfer: The framework enables the transfer of complex manipulation skills from a single kinesthetic demonstration to unseen objects within the same category, requiring no additional policy training.
Geometry-First Paradigm: Unlike end-to-end learning approaches, GIFT explicitly encodes task constraints via surface functions and screw geometry, ensuring interpretability and strict constraint fidelity.
Robustness to Shape and Topology Changes: By leveraging functional maps and screw interpolation, the method successfully handles objects with significantly different shapes and reference frames while maintaining the invariants of the original task.
Data Efficiency: The approach relies on geometric structure rather than large-scale datasets, making it computationally efficient and suitable for scenarios where data is limited.

Experimental Results
The authors validated GIFT on a 7-DoF KUKA iiwa robot with a Schunk PG 70 gripper and wrist-mounted RGB-D camera. Experiments included four skills: wiping, drawing a box, bottle stirring, and button pressing, tested across various object instances.

Object Matching: The system successfully identified the correct target object in multi-object scenes by selecting the object with the most diagonally dominant functional map.
Function Transfer: Mean Absolute Error (MAE) for transferred RIF and EIF functions remained low across all skills. The average processing time for function transfer was approximately 7.4 seconds.
Task Execution: The system achieved high success rates (75% to 100% depending on the task) in executing transferred skills in novel scenes. Trajectory constraints were maintained even when the final pose differed from the initial configuration.
Limitations Observed: In some cases (e.g., wiping a symmetric block), the FMC did not account for object symmetry, leading to potential misalignment or rotation of the object. However, the authors note this did not constitute a total failure of the skill transfer.

Significance and Claims
The paper claims that GIFT offers a fully orthogonal and complementary approach to data-driven foundation models. While acknowledging the rapid progress of VLA and diffusion policies, the authors argue that GIFT targets a specific regime where explicit constraint fidelity, explainability, and predictable generalization across substantial shape variations are paramount. The significance lies in providing a mechanism where constraint satisfaction follows directly from the representation and interpolation choices, rather than being inferred from data. The authors posit that this geometric grounding can eventually be layered beneath large-scale visuomotor policies to "harden" real-world behavior, bridging the gap between robust geometric grounding and semantic flexibility.