Next-Balance / README.md
Lamapi's picture
Publish Next Balance
4cefcfa verified
|
Raw History Blame Contribute Delete
3.57 kB
metadata
title: Next Balance
emoji: ⚖️
colorFrom: green
colorTo: gray
sdk: static
app_file: index.html
pinned: false
license: apache-2.0
short_description: Adaptive vs PID control under payload changes (simulation)

Next Balance

Next Balance keeps a simulated lift platform at 0.5 m while its payload changes at t≈4 s and friction rises at t≈7 s. It compares a fixed PID controller with an adaptive controller that estimates mass, viscous friction and Coulomb friction online (recursive least squares on filtered signals) and rescales its control law. Simulation only.

Model: https://huggingface.co/thelamapi/Next-Balance — set load change, actuator delay, sensor noise and friction; both controllers re-run in the browser.

Results on held-out scenarios (200 seeds per split, same disturbances for both controllers)

Split Controller Max deviation after load change, mean / p90 (m) Settling time, median (s) Settled Control energy (kN²·s)
nominal Fixed PID 0.144 / 0.283 1.97 81% 0.061
nominal Adaptive (Next Balance) 0.140 / 0.263 5.03 96% 0.052
stress Fixed PID 0.286 / 0.540 2.19 68% 0.098
stress Adaptive (Next Balance) 0.292 / 0.504 5.10 90% 0.094

Stress split by condition (mean max deviation in m, settled rate):

Condition n PID Adaptive PID settled Adaptive settled
heavy_load_change(dm>4kg) 121 0.322 0.406 73% 88%
load_removed(dm<0) 10 0.234 0.035 20% 100%
long_delay(>=40ms) 88 0.447 0.359 51% 82%
high_noise(>2mm) 143 0.294 0.310 65% 90%

Paired wins (adaptive better on the same scenario): nominal max deviation 86/200, stress max deviation 69/200, stress energy 148/200. Settling time is measured from the load change until the platform last leaves a ±1 cm band (the later friction change counts). Runs that leave the band within the final 0.5 s count as not settled.

Two held-out stress scenarios

Replay video: the first held-out stress scenario and the stress scenario with the adaptive controller's largest deviation.

Method

  • Plant: m(z̈+g) = u(t−delay) − c·ż − F·tanh(ż/0.01), dt = 10 ms, force limited to [0, 400] N, platform 2.0 kg.
  • Shared α-β estimator (α=0.3, β=0.05). Both controllers only know the nominal payload.
  • Fixed PID: Kp=300, Ki=400, Kd=80 plus nominal gravity feed-forward.
  • Adaptive: RLS (λ=0.99) on [m, c, F] with state-variable filters (ω=10.0 rad/s); control u = m̂(ωn²e − 2ζωn v + ki∫e + g) + ĉv (F̂ is estimated but not fed back because it chatters with noisy velocity), ωn=6, ζ=0.7, ki=30. The delay value is given to the controller to align its regression.
  • The 0.5 m target is approached with a 1 s ramp starting at t = 0.5 s.
  • Both gain sets come from grid searches of the same size and cost (IAE + 0.5·energy) on 40 tuning seeds; test seeds are disjoint.
  • The browser runs a line-by-line JavaScript port; Node parity on 6 fixed runs: max abs difference 8.459899447643693e-14.

Usage

python inference.py --dm 8 --delay 4 --noise 0.003   # prints metrics for both controllers
python inference.py --selftest

Limitations

  • 1-D rigid-body model; no structural flexibility, actuator dynamics beyond pure delay, or sensor bias.
  • The adaptive controller is told the actuator delay; an unknown delay would need estimation.
  • Results are from simulation on NVIDIA L4; no physical platform was tested.