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

TL;DR BoBDeLaN replaces the LSTM context encoder of a payload-adaptive Deep Lagrangian MPC with a low-pass filter and a recurrent network whose sensitivity to disturbances is bounded by construction. On a real Franka pick-and-place with injected 5 and 10 Hz disturbances, its velocity error grows only from 88% to 116% of nominal MPC, while the LSTM version's reaches 665%.
What they do
Context-Aware DeLaN learns a Lagrangian residual model of a robot arm, conditioned on a latent context inferred from recent measurements, so that MPC can adapt to an unknown payload. Nothing limits how much noise or disturbances in that history move its LSTM-inferred context. BoBDeLaN swaps the LSTM for a recurrent network with built-in sensitivity bounds, preceded by a low-pass filter, so slow payload changes pass through and fast disturbances are attenuated.
How it works
- Encoder. The history of joint states and residual torques (relative to a nominal model) is normalized per channel, passed through two cascaded first-order low-pass filters (2 Hz), then into a Recurrent Equilibrium Network (REN) parametrized to be contracting with incremental gain at most 0.5 throughout training.
- Guarantees. Contraction makes the encoder forget its initial state. The gain bound and the filter bound, frequency by frequency, how far history perturbations can move the context, and hence the residual torque predictions if the DeLaN is Lipschitz in the context.
- Control. Nominal model plus context-conditioned DeLaN form the dynamics of a 50 Hz nonlinear MPC on a 7-DoF Franka Panda. Models are trained on 2020 simulated trajectories covering 100 payloads of 0-4 kg.
Results
- Real-robot data (12 runs, 0.2-24 Hz disturbances): REN context sensitivity falls off at high frequency; the LSTM’s peaks near 5 Hz despite the same input filter.
- Simulation, 600 runs per method with unseen in-distribution payloads: position error is 45% (REN) vs 28% (LSTM) of nominal MPC, residual torque error 53% vs 64%, acceleration RMS 1.8x vs 2.7x nominal. With 5-8 kg payloads, torque error is 39% vs 48%.
- Real Franka pick-and-place picking up 3 kg and 2 kg dumbbells: torque error 33% (REN) vs 62% (LSTM) of nominal, acceleration RMS about 140% vs 290%; the LSTM tracks position better (45% vs 78%). With disturbances, the LSTM’s velocity and acceleration errors reach 665% and 1680% of nominal.
Limitations
The bounds cover the encoder and, under a Lipschitz assumption on the DeLaN that the REN does not enforce, torque predictions along a fixed trajectory; they give no closed-loop tracking or stability guarantee. Stronger filtering also slows the response to real changes. The LSTM achieves lower position error on hardware and with in-distribution payloads in simulation.
Videos
Figures



Abstract
Learned physics-based models provide structure for robot control, but uncertain payloads and unmodeled dynamics limit their predictive accuracy. Context-aware Deep Lagrangian Networks enable adaptation through a latent context inferred from measurement histories while retaining Lagrangian structure for each fixed context. However, this structure alone does not constrain sensitivity to measurement disturbances. We introduce BoBDeLaN, which combines a contracting, gain-bounded Recurrent Equilibrium Network context encoder with stable linear filtering to shape disturbance sensitivity across frequencies. This architecture guarantees exponential forgetting of the recurrent initialization and admits a finite-horizon, frequency-weighted bound on context perturbations, which extends to residual torque prediction under a Lipschitz continuity assumption on the context-conditioned dynamics model. We evaluate a low-pass realization within model predictive control on simulated and real hardware manipulation tasks, together with frequency-sensitivity analysis on real-robot data. Compared with an LSTM encoder using the same input filter, BoBDeLaN exhibits reduced high-frequency context sensitivity, lower residual torque prediction error, and smoother motion, with less degradation under injected disturbances.
Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.
