Seonglae ChoHow to simulate via a sensory interface
(a) Retina mapping
- MuJoCo camera → downsample
- convert to photoreceptor neuron firing
(b) Olfaction (smell)
- No real chemical sensing so instead:
- an "odor scalar field"
- mapped to activation of a specific neuron group
(c) Touch
- contact force → mechanosensory neuron firing
- motor neuron firing → joint torque
Example:
- a specific neuron cluster → leg movement
Here too:
hard-coded mappings are almost required
4) Calibration / tuning
This is the key trick:
- synaptic weight scaling
- global gain tuning
- noise injection
- homeostasis constraints
5) (almost certainly) some learning / fitting is involved
A pure copy still won’t run, so:
- RL fine-tuning
- parameter fitting
- or:
- behavior cloning (partially)
In other words: it’s not reproducing the senses; it’s translating sensory signals into inputs compatible with the sensory neurons.
MujoCo model:
flybody
TuragaLab • Updated 2026 Apr 15 14:1
volume
