Recognition of Manipulated Objects by Motor Learning

Hiroaki Gomi, Mitsuo KawatoNeurIPS 1991

This classic paper from the early 1990s NIPS era explores a fundamental challenge in robotics: how can neural networks learn to recognize and manipulate different objects by seamlessly integrating sensory feedback and motor commands?

By employing a technique known as "feedback-error-learning," the authors demonstrate how robotic systems can identify objects through both visual cues and "somatic" (proprioceptive) feedback. This allows a robot to dynamically adapt to varying physical properties—like mass and friction—without requiring a pre-defined, perfect teacher.

The Problem: The "One-Size-Fits-All" Failure

In robotics and motor control, a static controller fails the moment the environment changes. If a robotic arm picks up a heavy hammer and then switches to a light screwdriver, the physics (inverse dynamics) required to move the arm change drastically.

Conventional feedforward neural networks struggle with this dynamic environment because they cannot instantly switch their internal control laws. Furthermore, in real-world learning scenarios, we rarely have access to the "exact target motor command" needed to train a network. Often, we only know if we missed our target. The authors aim to bridge this gap by creating a system that recognizes an object through the act of manipulation and adjusts its control strategy on the fly.

Methods: Feedback-Error-Learning

The researchers introduce two distinct modular neural network architectures based on Feedback-Error-Learning (FEL). In FEL, a simple, conventional feedback controller acts as a "teacher" for a more sophisticated neural network. The feedback signal itself serves as the error signal to train the feedforward component.

Scheme 1: The Gating Network (Competitive)

Based on the "Mixture of Experts" architecture, this scheme uses a competitive approach to select the right control law for discrete objects.

Architecture of the Gating Network scheme, illustrating the integration of visual and somatic information to select appropriate expert networks.
  • A Gating Network receives somatic (position/velocity) and/or visual information.
  • It acts as a switch, selecting which Expert Network (each representing a specific object's dynamics) should be active at any given moment.
  • The Innovation: The system uses a "quasi-target vector" ($u' = u + u_{fb}$), where the training signal is the sum of the current motor command and the feedback correction, allowing the network to learn without an explicit external teacher.

Scheme 2: The Estimation Network (Cooperative)

Instead of picking a discrete expert, an Estimation Network predicts continuous physical parameters (like mass or friction). These estimated parameters are then fed into a functional controller that adjusts its output accordingly. This cooperative approach allows the system to handle an infinite variety of objects within a specific continuous domain (e.g., any weight between 1kg and 5kg).

Experimental Setup

As a foundational computational paper, the authors validated their architectures using simulated environments:

  • Physical Model: A 1D mass-spring-damper system with varying Mass ($M$), Viscosity ($B$), and Stiffness ($K$).
  • Visual Cues: A simple $4 \times 4$ retinal matrix where distinct patterns represented different objects.
  • Trajectories: Desired movement paths were generated using an Ornstein-Uhlenbeck random process to simulate natural, unpredictable motion.

Key Insights

  • Somatic Recognition: We don't just recognize objects by looking at them. The "feel" of an object—how it resists or assists motion—is a primary data source for identification. The network successfully uses this somatic feedback to classify objects.
  • Functional vs. Declarative Representations: The networks do not merely learn a declarative label (e.g., "This is Object A"). They learn a functional representation—the specific inverse dynamics required to physically manipulate that object.
  • Vision as a Predictor: While somatic information is sufficient for recognition during motion, visual information allows for predictive identification before the object is even moved, resulting in significantly smoother initial control.
  • The "Identical Box" Problem: By combining vision and somatic data, the system can distinguish between two objects that look identical but have different weights—a task impossible using vision alone.

Key Results

The simulations yielded strong evidence for the efficacy of these modular networks:

Performance of the Gating Network showing successful switching between expert networks (Gate 1, 2, 3) as different objects are manipulated.
  • Successful Adaptation: The Gating Network correctly identified objects (labeled $\alpha, \beta, \gamma$) and switched to the appropriate expert network within milliseconds of manipulation.
  • Error Reduction: After the learning phase, the feedback motor command ($u_{fb}$) dropped to near zero. This indicates that the feedforward neural network had successfully internalized the dynamics and taken over the task with high precision.
  • Generalization: When presented with an "unknown" object with parameters falling between the learned experts, the Gating Network selected the closest known expert, providing a reliable "best-guess" control law.
  • Continuous Tracking: The Estimation Network successfully tracked a continuously changing mass in real-time, allowing the robotic trajectory to stay on target despite shifting physics.

Limitations and Future Directions

While groundbreaking, the authors noted a few limitations. Using "quasi-targets" (feedback error) for training is approximately five times slower than training with exact, known motor targets. Additionally, while the Estimation Network works exceptionally well for a single parameter like mass, estimating multiple complex, non-linear physical parameters simultaneously remains challenging.

Finally, the Gating Network approach is inherently limited by the number of experts. Scaling to thousands of distinct objects would require an impractical number of expert networks. The authors suggest that future robust systems should integrate both schemes: using gating networks for broad, discrete categories, and estimation networks for continuous fine-tuning within those categories.

Summary updated

BibTeX
@inproceedings{gomi1991recognition,
  author = {Hiroaki Gomi and Mitsuo Kawato},
  title = {{Recognition of Manipulated Objects by Motor Learning}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {1991},
}
Figure