Readout adaptation in simulated arm reaching: learned compensation and performance after updates stop
- 1Kyushu Institute of Technology
Abstract
When an arm learns to compensate for a force that deflects its movement, removing the force can produce an opposite deviation, or after-effect. We evaluate what the controller retains by measuring the after-effect, the learned torque correction, and tracking performance after weight updates stop. We use a simulated torque-driven planar two-link arm reaching in eight directions under velocity-dependent sideways forces. Only an external linear torque readout adapts online; the parameters generating its features stay fixed. Features come from three configurations of closed-form continuous-time (CfC) networks, an echo-state network, or a radial-basis-function layer. All three CfC configurations, including neural circuit policy wiring, are trained for 400 epochs. These five feature-based variants use least-mean-squares (LMS) learning; a sixth uses recursive least squares (RLS) on the dense-backbone CfC readout. Learning rates are selected on three model initializations and evaluated on ten separate initializations under three force fields. All six variants improve tracking relative to their fixed-readout baselines and pass our after-effect criterion in every evaluation run. Diagnostics on the same ten evaluation initializations in the main field (60 conditions) quantify differences behind this shared decision: the median normalized squared residual between the learned output change and ideal compensation ranges from 0.0097 to 0.383 across implementations on common prescribed inputs. With updates stopped and feedback control active, all final readouts retain useful compensation, reducing tracking error by 40-97% relative to their pre-adaptation versions. Compared with the final adaptive cycle, the five LMS variants have median fixed-readout error ratios of 2.28-5.87, while RLS remains nearly unchanged. These measurements distinguish the performance achieved during adaptation from that retained in the final weights. They provide a concrete evaluation procedure for readout adaptation: measure the learned output correction and test the final controller with updates off, alongside behavioral after-effects. The comparison characterizes complete implementations in this task; differences in preparation and hyperparameters limit attribution to feature representations or update rules alone.
Related
- Vibrotactile stimulation on the upper back requires stronger vibration during goal-directed reaching but does not impede reaching performance
- Effects of feed intake-regulated nutritional compensation strategies on growth performance, blood indices, and rumen microbiota of Gangba sheep
- Effect of a physical training program on simulated laparoscopic performance in gynecologic surgery: a randomized controlled trial
- Aging and Hearing Loss in Emotional Prosody Processing: Divergent Neural Adaptation and Behavioral Performance
- Beyond conflict adaptation: Affectively experiencing and observing conflict enhance performance across conflict and non-conflict trials
BibTeX
@misc{kobayashi2026readout,
title = {{Readout adaptation in simulated arm reaching: learned compensation and performance after updates stop}},
author = {J. Kobayashi},
howpublished = {bioRxiv},
year = {2026},
month = jun,
doi = {10.64898/2026.06.11.731593},
url = {https://www.biorxiv.org/content/10.64898/2026.06.11.731593v4},
note = {CC BY},
}
RIS
TY - UNPB
TI - Readout adaptation in simulated arm reaching: learned compensation and performance after updates stop
AU - J. Kobayashi
PY - 2026
DA - 2026/06/11
PB - bioRxiv
DO - 10.64898/2026.06.11.731593
UR - https://www.biorxiv.org/content/10.64898/2026.06.11.731593v4
AB - When an arm learns to compensate for a force that deflects its movement, removing the force can produce an opposite deviation, or after-effect. We evaluate what the controller retains by measuring the after-effect, the learned torque correction, and tracking performance after weight updates stop. We use a simulated torque-driven planar two-link arm reaching in eight directions under velocity-dependent sideways forces. Only an external linear torque readout adapts online; the parameters generating its features stay fixed. Features come from three configurations of closed-form continuous-time (CfC) networks, an echo-state network, or a radial-basis-function layer. All three CfC configurations, including neural circuit policy wiring, are trained for 400 epochs. These five feature-based variants use least-mean-squares (LMS) learning; a sixth uses recursive least squares (RLS) on the dense-backbone CfC readout. Learning rates are selected on three model initializations and evaluated on ten separate initializations under three force fields. All six variants improve tracking relative to their fixed-readout baselines and pass our after-effect criterion in every evaluation run. Diagnostics on the same ten evaluation initializations in the main field (60 conditions) quantify differences behind this shared decision: the median normalized squared residual between the learned output change and ideal compensation ranges from 0.0097 to 0.383 across implementations on common prescribed inputs. With updates stopped and feedback control active, all final readouts retain useful compensation, reducing tracking error by 40-97% relative to their pre-adaptation versions. Compared with the final adaptive cycle, the five LMS variants have median fixed-readout error ratios of 2.28-5.87, while RLS remains nearly unchanged. These measurements distinguish the performance achieved during adaptation from that retained in the final weights. They provide a concrete evaluation procedure for readout adaptation: measure the learned output correction and test the final controller with updates off, alongside behavioral after-effects. The comparison characterizes complete implementations in this task; differences in preparation and hyperparameters limit attribution to feature representations or update rules alone.
ER -