Can a robot learn from its own lucky accidents?
A neural robot controller learned new behaviours, like turning a particular way when facing a mirror, simply by being retrained on runs where it had happened to do the right thing.
Source
Serendipitous Offline Learning in a Neuromorphic Robot
Study at a glance
- Design
- Computational / modelling — Spiking-network controller on SpiNNaker hardware driving a tracked robot in a T-maze, plus abstract simulations varying signal strength and number of positive examples.
- N
- No single N; robot tasks were trained from ten positive examples each, and simulation points were averaged over repeated random trials.
- Population
- A small mobile robot controlled by a neural network built with the Neural Engineering Framework, and simulated sensory signals.
- Outcome
- Similarity between network output and the desired action (normalised dot product); robot turning behaviour in a T-maze with and without a mirror.
Structured fields used in claim comparison tables when every cited study has a complete layer.
What they did
The authors hand-built simple reflexes (drive forward, avoid walls, turn away from obstacles) as a spiking neural network running on low-power neuromorphic hardware. They let the robot wander a T-maze, kept the recordings of runs where it happened to do the desired thing, and trained new sensor-to-action connections to reproduce those runs. They also simulated an abstract version in which a weak or strong signal was buried in random sensory noise, varying how many positive examples were used.
What they found
In the idealised simulation, output matched the training example with a similarity of 0.99. With strong signals, a few examples were enough to generalise, while with weak signals the network latched onto chance correlations and improved only as examples accumulated, even when the signal was absent. On the real robot, ten positive examples were enough to make it turn left consistently, and to turn right when a mirror was present and left when it was not.
The limits
What it doesn't show
The robot results are demonstrations with plotted trajectories rather than systematic success rates across many tasks, environments or robots. The authors say it is not yet clear which real tasks provide strong enough sensory signals. The method uses only positive examples, with no punishment for wrong actions, and is not compared against standard reinforcement or imitation learning baselines.
Key terms
- Neuromorphic hardware
- Chips designed to run large networks of spiking neurons efficiently and with low power, such as SpiNNaker.
- Positive examples
- Recorded episodes where the robot happened to perform the desired behaviour, used as training targets.
- Spurious correlation
- A chance relationship in a small sample that does not reflect a real link, which can mislead a learner.
- Overfitting
- Fitting the quirks of a few training examples so closely that performance on new situations suffers.
- Neural Engineering Framework
- A method for choosing connection weights so that populations of spiking neurons approximate chosen functions.
Flashcards
0 of 10 answers reviewed
Research intelligence for this paper
See its role on concept claims, tensions it is part of, placement history, and related discoveries.
Quiz yourself
What is 'serendipitous' about the learning method?
Common questions
How is this different from ordinary reinforcement learning?
There is no reward signal during behaviour; a person picks out successful runs afterwards and the network is trained to reproduce them, using positive examples only.
Why did performance improve with more examples even when the signal was zero?
With few examples the network used chance correlations in the noise, a kind of superstition; more examples averaged these away so it settled on a sensible constant output.
Does this prove the robot 'recognises' mirrors?
It shows the network found some sensory difference caused by the mirror and linked it to a turn, but not what feature it used or how robust that is.
More on Robot learning