Polymander: a salamander robot that swims and walks

Year: 2026
Context: EPFL · Computational Motor Control course
Team: Team of 3
Code: on request

TL;DR

Problem: Produce efficient, robust swimming and walking from coupled oscillators driven by a single descending drive signal.
What we built: In stages: an open-loop traveling wave, a 16-oscillator double-chain CPG tuned to X-ray recordings of real salamanders, stretch feedback, then contact-driven switching between swimming and walking.
Result: +20% swimming speed and −16% cost of transport with feedback, 61% lower joint error against the animal data, and 2.4× more distance under neural damage.
Stack: Python, NumPy/SciPy, FARMS (MuJoCo), Matplotlib

How much does a little sensory feedback help a spinal-cord-like controller, especially when parts of it fail?

+20% SPEED · −16% CoT
CPG
Sensory feedback
MuJoCo (FARMS)
Bio-inspired control

Top view of the simulated Polymander salamander robot swimming with a traveling body wave

Watch clip

The approach

Baseline. An open-loop traveling wave drives Ekeberg muscle models; a 5 × 5 sweep of amplitude and wave lag maps out what works.
Spinal CPG. A double chain of 16 amplitude-phase oscillators, swept over drive and phase lag. Steering comes from left/right drive asymmetry.
Match the animal. We compared joint trajectories with X-ray recordings of swimming salamanders, then retuned the network to match their uneven phase lag along the body.
Feedback and damage. Proprioceptive stretch feedback with a gain sweep, then a robustness study with 0–20% random disruption of sensors and couplings.
Amphibious switching. A body and limb CPG where ground contact pulls the drive toward walking (d = 2.5) and its absence toward swimming (d = 4.5).

What it achieved

+20% speed, −16% cost of transport. Stretch feedback (w_ipsi = 3) raised swimming speed from 0.201 to 0.241 m/s and cut cost of transport by 16% at the same drive.
61% closer to the animal. Tuning to the X-ray data cut mean joint-trajectory error from 0.236 to 0.092 rad.
2.4× more travel under damage. With 20% random disruption, the decoupled, sensory-driven network kept about 2.4 times more forward travel than the combined (coupled) network.
Steering. Asymmetric drive turns the robot, with curvatures of about +16 and −20 m⁻¹.
Swim-to-walk and walk-to-swim. Both transitions happen in an amphibious arena, triggered by ground contact alone.

Footage & figures

Walking (video still) — Clip 01Walking: the limb oscillators take over at low drive.

Swim to walk (video still) — Clip 02Swim to walk: when the feet touch the ground, the drive drops into the walking range.

Walk to swim (video still) — Clip 03Walk to swim: without ground contact, the drive rises into the swimming range.

Fig. 01Stretch-feedback gain sweep. As the gain grows, swimming gets faster and cheaper, and the network's rhythm speeds up.Team figure · course reportFull size image: Swimming speed, cost of transport, neural frequency and amplitude as the stretch-feedback gain increases (opens in a new tab)

Fig. 02Robustness to 20% random disruption of sensors and couplings. The combined (coupled) network nearly stops; the decoupled, sensory-driven network keeps swimming at a far lower cost of transport.Team figure · course reportFull size image: Speed and cost of transport for the combined (coupled) and decoupled networks, with and without 20% neural disruption (opens in a new tab)

Fig. 03Forces on the body across the drive range: ground reaction dominates at walking drives, and hydrodynamic force takes over past the walk/swim cut-off at d = 3.Team figureFull size image: Ground reaction and hydrodynamic forces on the body as the drive moves from walking to swimming (opens in a new tab)

Notes

The Polymander robot model, the simulation setup and the salamander X-ray recordings were provided by the course. The code lives in a private team repository.