BoBDeLaN paper adds bounded-bandwidth context adaptation for robust deep Lagrangian robot control
Jan_R_Peters · x · 2026-11-21
Researchers from TU Darmstadt (Lucas Schulze, Jan Peters, Oleg Arenz, et al.) released BoBDeLaN, targeting the accuracy loss of physics-based learned models in robot control caused by uncertain payloads and unmodeled dynamics.
Key ideas:
- Builds on context-aware Deep Lagrangian Networks, adding a contracting, gain-bounded Recurrent Equilibrium Network context encoder plus stable linear filtering to shape disturbance sensitivity across frequencies.
- Guarantees exponential forgetting of the recurrent initialization and provides a finite-horizon, frequency-weighted bound on context perturbations, extendable to residual torque prediction under a Lipschitz assumption.
In MPC experiments on simulated and real manipulation tasks, BoBDeLaN shows lower high-frequency context sensitivity, smaller residual torque prediction error, and smoother motion than an LSTM encoder with the same filter, degrading less under injected disturbances. Submitted to L4DC 2027.
More from Embodied
- Huawei's clip-on Ruiying Z10 brings 10x optical zoom, journalists ditch cameras — CodeByPoonam · 2026-11-21
- Refining the morphogenesis-to-locomotion pipeline for automated robot design — zzznah · 2026-11-21
- Uncertainty-Aware Robotic World Model Talk Draws Crowd at CoRL 2026 Workshop — breadli428 · 2026-11-21
- BoBDeLaN: Combining Deep Lagrangian Networks and Stable Filters for Robust Robot Control — Jan_R_Peters · 2026-11-21
- Teleoperating an OpenYAM robot arm with AR hand tracking and an IK solver — AjdDavison · 2026-11-21
- ACE Robotics humanoids run the night shift at a Shanghai hotel, handling laundry and deliveries — Scobleizer · 2026-11-21