Reproduce “Training slow silicon neurons to control extremely fast robots with spiking reinforcement learning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Irene Ambrosini, Ingo Blakowski, Dmitrii Zendrikov, Cristiano Capone, Luna Gava, Giacomo Indiveri, Chiara De Luca, Chiara Bartolozzi
The authors train a spiking-neuron system to intercept fast-moving pucks in a simulated air-hockey task, with a DYNAP-SE neuromorphic chip in the loop. Its fixed, randomly connected network acts as a reservoir, while reward-guided local updates train a small action readout. With randomized puck position and speed, success stabilizes above 98% after 1,300 episodes; this is not a physical-robot result.
Key insights
- Neuromorphic systems can effectively perform predictive control for fast robotic manipulation tasks with tight temporal constraints, moving beyond simpler benchmarks.
- Combining fixed random connectivity in a recurrent SNN (reservoir computing) with a local e-prop learning rule enables efficient online learning of control policies directly on neuromorphic hardware.
- Exposing the learning agent to broad state distributions (randomized puck positions and speeds) improves generalization and leads to more robust, coherent closed-loop control strategies.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.