Gain-bounded REN replaces LSTM encoder: adaptive robot control stays smooth under disturbance

BoBDeLaN: Bounded-Bandwidth Context Adaptation for Deep Lagrangian Robot Control

Lucas Schulze, Sucheth Shenoy, Jan Peters, Oleg Arenz

cs.RO, cs.LG, eess.SY

2026-11-16

Swapping DeLaN's LSTM context encoder for a gain-bounded REN behind a low-pass filter pays off under disturbance: velocity error rose to 116% of nominal vs 665% for the LSTM.

What problem this solves

A robot arm that picks up an unknown payload changes its own dynamics on the spot, and a fixed learned model drifts out of range. Context-Aware DeLaN (an earlier paper from the same group) handles this with an LSTM encoder that reads a short history of joint states and residual torques, infers a latent context z, and conditions a Lagrangian dynamics model on it. The model adapts to whatever the arm is holding while staying mechanically consistent for each fixed z, and plugs into MPC for trajectory tracking.

The encoder is the weak link. LSTM training constrains neither how fast the initial hidden state is forgotten nor how far the inferred z drifts when the measurement history is perturbed. Sensor noise and estimation errors enter the history, shift z, and drag the dynamics predictions supplied to MPC along with them; low validation error says nothing about either behavior. The paper states two requirements: exponential forgetting of the initial state (contraction) and bounded output sensitivity to history perturbations. Standard recurrent architectures guarantee neither.

Method

BoBDeLaN swaps the LSTM for a Recurrent Equilibrium Network (REN), a recurrent family whose parametrization bakes stability properties in, preceded by two cascaded first-order low-pass filters set at fb = 2 Hz.

Everything still runs inside CaDeLaC: before each control update the encoder re-estimates z, holds it fixed across the MPC prediction horizon, and the optimizer trades tracking error against torque effort.

Results

On recorded Franka Panda data (12 runs) with 0.2–24 Hz sinusoidal injections at half a standard deviation per channel, REN context and torque responses attenuate with frequency while the LSTM peaks near 5 Hz.

SettingMetric (% of nominal MPC)RENLSTM
Sim, in-distribution payloads, 600 runsposition error45%28%
Sim, in-distributionresidual torque error53%64%
Sim, in-distributionacceleration RMS1.8x2.7x
Hardware, no disturbancevelocity error88%149%
Hardware, no disturbanceresidual torque error33%62%
Hardware, with disturbancevelocity error116%665%
Hardware, with disturbanceacceleration error140%1680%

The hardware task is a 26 s pick-and-place with four dynamics changes (a 3 kg and then a 2 kg dumbbell) and 5/10 Hz injections on the first five joints. The trade is consistent everywhere: the LSTM posts the lower raw position error (45% vs 78% undisturbed), while the REN wins velocity error, torque error, smoothness, and every disturbed condition. On out-of-distribution payloads (5–8 kg against a 0–4 kg training range) both reach 33% position error, with the REN ahead on torque error (39% vs 48%) and torque-rate RMS (328% vs 455%).

Why it matters

This is a replicable recipe for attaching control-theoretic guarantees to a learned system: swap one encoder module and you get initialization forgetting, a bounded gain, and frequency shaping at once, with the training loop untouched. The cost shows up as encoder expressiveness; the undisturbed position error on hardware is 33 percentage points worse than the LSTM. The fit is manipulation with noisy sensing and frequent disturbances, where smooth control matters more than peak tracking accuracy. Incremental as an improvement, but the direction, certified sensitivity bounds for adaptive dynamics learning, had essentially no prior takers.

Limitations

Stated in the paper:

Two more from reading it: the low-pass choice filters out fast but informative effects such as friction and actuator dynamics along with the noise, and the authors list frequency-selective filters as future work. The comparison pool is also thin, only the same-filter LSTM and nominal MPC, with no online model-update baselines.

Terms

Source

What people are saying

Related papers

All paper explainers