IROS 2026 FILIC - Heungwoo/research GitHub Wiki

IROS 2026 — FILIC: Dual-Loop Force-Guided Imitation Learning with Impedance Torque Control

Venue: IROS 2026 (Pittsburgh) · paper #1885 · Tsinghua University · HKUST(GZ) · DISCOVER Robotics (Ge, Jia, Li, … Zhou). Paper: arXiv 2509.17053 · code. Representative of: force-aware learning from demonstration — a dual-loop framework that makes a position-centric IL policy force-informed and force-executed, even on arms without F/T sensors. Companions: Tactile VLA · Real-Time Execution · Dexterous Manipulation · IROS 2026 survey.

FILIC's dual-loop structure — outer loop: two-view ResNet features + an external force estimator (from joint torque τ) feed, via cross-attention, a Transformer encoder-decoder that outputs a 25 Hz action pose sequence; inner loop: an impedance controller (2 kHz, virtual spring-damper) + gravity compensation (250 Hz) execute compliantly through IK (architecture figure from Ge et al., arXiv 2509.17053, © the authors)

1. Problem

Many contact-rich tasks need precise force regulation, but most imitation-learning policies are position-centric and force-unaware, and adding F/T sensors to collaborative arms is costly + extra hardware.

2. Method

FILIC = Force-guided Imitation Learning + Impedance torque Control, in a dual-loop structure:

  • A Transformer-based IL policy paired with an impedance controller → compliant, force-informed, force-executed manipulation.
  • For arms without F/T sensors: a cost-effective end-effector force estimator from joint-torque measurements via analytical Jacobian inversion, compensated with model-predicted torques from a digital twin.

3. Results

  • FILIC significantly outperforms vision-only and joint-torque-based methods, achieving safer, more compliant, adaptable contact-rich manipulation. Code released.

4. Why it matters

FILIC is IROS 2026's force-from-demonstration representative and a pragmatic answer to a recurring gap (data-pyramid §4: force/tactile is the signal lost first): it makes IL force-aware without new sensors by estimating EE force from joint torque + a digital-twin correction. The dual-loop (learned intent + impedance execution) is the transferable idea — the control-theoretic counterpart to the multi-rate tactile loops in MrGrasp/T-Rex: instead of adding a tactile sensor stream, it recovers force analytically and executes compliantly. This lowers the hardware bar for contact-rich LfD.

Limitations (reviewer): joint-torque force estimation is coarser than a real F/T sensor (bounded by digital-twin fidelity); impedance control assumes a compliant/torque-controllable arm; contact-rich-scoped.

5. Links

← Back to IROS 2026 survey · Home