The Salamander Robot

Salamanders occupy an interesting position. They swim like fish, using a travelling wave down the body. They also walk on land with a stepping gait. And they are close to the point in vertebrate history where limbs appeared.

That raises a structural question about how the nervous system supports both. Two hypotheses are available.

Either the transition to land required a new locomotor circuit, with the swimming machinery replaced or set aside, or the primitive swimming circuit was kept and something was added on top of it.

The second is more parsimonious and harder to test, since you cannot easily observe the internal organisation of a spinal cord in a behaving animal.

The approach taken was to build the hypothesis into hardware. Auke Ijspeert and colleagues published "From Swimming to Walking with a Salamander Robot Driven by a Spinal Cord Model" in Science, in the issue of 9 March 2007, work carried out at EPFL with the INSERM research centre and the University of Bordeaux.

The paper presents a spinal cord model and its implementation in an amphibious salamander robot, demonstrating how a primitive neural circuit for swimming can be extended by phylogenetically more recent limb oscillatory centres.

The model has two populations of oscillators, and the relationship between them carries the whole argument.

The body CPG is a chain of oscillators along the trunk and tail, coupled to neighbours with a phase lag, exactly the travelling-wave arrangement from the previous lesson. This is the evolutionarily older circuit, the one a lamprey has.

The limb CPG is a small set of oscillators, one per limb, coupled to each other to produce the phase relationships of a walking gait. This is the newer addition.

The critical detail is the connection between them. The limb oscillators project onto the body oscillators, and when active they impose their own slower rhythm on the body chain.

That single asymmetric coupling produces the behavioural switch.

When the limb oscillators are silent, the body chain runs freely at its own frequency with its built-in phase lag, giving a travelling wave down the body. The robot swims.

When the limb oscillators are active, they dominate the body chain, forcing it into a standing wave synchronised with the stepping cycle, with the trunk bending to extend each step. The robot walks.

One circuit, two behaviours, selected by whether an added module is running.

The property that makes the result striking is how little information the high-level controller sends.

In the experiments, simple electrical signals of the kind the brain sends to the spinal cord were transmitted wirelessly from a laptop to the robot, and those signals were sufficient to change its speed and direction and to switch it from walking to swimming.

The mechanism is a drive parameter, a single scalar representing descending activation.

The two oscillator populations respond to drive differently, and that difference is the entire design.

The limb oscillators are active only in a low range of drive. Below a threshold they are silent; above an upper threshold they saturate and stop oscillating.

The body oscillators remain active across the whole range, increasing frequency as drive rises.

So increasing drive from zero walks the robot faster and faster, until drive passes the limb oscillators' upper limit. At that point the limbs stop oscillating and fold back, the body chain is released from their influence, and the travelling wave reappears. The robot swims.

Gait transition as a saturation phenomenon, not a decision, a mode switch, or a planner selecting a controller.

The transition is a consequence of two oscillator populations having different operating ranges. Nothing detects the water or decides to switch; a monotonically increasing scalar produces a qualitative change in behaviour when it crosses a threshold.

The full control interface reduces to a small set of parameters, and each maps cleanly onto a behavioural variable.

drive (scalar)
  low        limbs active, robot walks
  increasing walking speed rises
  high       limbs saturate, robot swims
  higher     swimming speed rises

drive asymmetry (left versus right)
  larger amplitude on one side turns the robot

body-limb coupling strength
  how strongly the standing wave locks to the step cycle

There is no separate steering controller, no gait selector, and no trajectory generator. Direction is asymmetry in the drive; speed is drive magnitude; gait is which side of a saturation threshold the drive sits on.

The paper's own framing is that the model suggests neural mechanisms for modulation of velocity, direction, and type of gait that are relevant for all tetrapods, and that the distributed neural system in the spinal cord holds the key to the complex locomotor capabilities of vertebrates.

That is a claim about biology, made using a robot as the evidence, which is what makes this work unusual rather than merely an engineering demonstration.

It is worth separating the claims carefully, because a robot working does not automatically establish a fact about animals.

Sufficiency. The model shows that a body CPG plus limb oscillators plus a single drive signal is enough to produce both salamander gaits and the transition between them. Before the demonstration, that was a plausible story; afterwards it was a demonstrated mechanism.

Consistency with evolution. The architecture supports the hypothesis that limb oscillators were added to an existing swimming circuit rather than replacing it, since the arrangement that works is exactly that one.

A mechanism for gait transitions. Saturation of one oscillator population with rising drive is a concrete proposal for why animals switch gaits at particular speeds.

And the limits, stated plainly. Sufficiency is not necessity. Showing this architecture works does not prove the salamander uses it; another organisation might produce the same behaviour. The robot is not a salamander, and its actuators, body compliance, and hydrodynamics differ.

The correct reading is that the robot converts a verbal hypothesis into a tested mechanism, which narrows the space of plausible explanations without closing it.

The 2007 model commanded joint angles directly. Later work in the same line made the model substantially more biological, and the additions matter because they change what the system can do.

Subsequent work reproduced multiple motor behaviours in a salamander robot using virtual muscles and a distributed CPG controller regulated by drive signals and proprioceptive feedback.

Three changes are involved.

Virtual muscles replace direct position commands with a model of muscle force generation, so the limb becomes compliant. It yields when it meets an obstacle instead of forcing through it, which is both safer and closer to how animals interact with terrain.

Proprioceptive feedback lets the actual state of the body influence the oscillators, so if a limb is delayed by rough ground, the rhythm adapts rather than continuing regardless.

Distributed control places oscillators throughout the body rather than centralising them, matching the biology and improving fault tolerance.

The unifying theme is the interplay between feedforward and feedback, which is what makes CPG-based control robust in the real world rather than only in simulation, and it is where the next lesson begins.

Collecting the results, since each one is a specific claim rather than a general endorsement.

The general engineering lesson is the one worth carrying beyond salamanders.

If you build the right dynamical structure, complex behaviour becomes controllable through very few parameters. The complexity lives in the coupled dynamics, which run autonomously; the interface exposes only what a higher level actually needs to decide.

That is a rather different design philosophy from computing what every joint should do at every instant, and the next lesson examines where it holds up and where modern systems supplement it.