Matt Strong

Matt Strong

CS PhD at Stanford University

← Paper Digest

ILA: A Transferable Interface-Level Adapter for Compliant Humanoid Motion Tracking

ZhuoChang Cai, Zihao Yan, Jinxuan Zhu, Yanjie Li, Hao Xiong

November 2026

TL;DR ILA is a small adapter, trained in simulation, that edits the arm reference a frozen humanoid motion tracker follows, so the robot yields when a person guides its hand. The same adapter cuts guidance-target error by 63.6-75.4% across three trackers and lets each of them complete wrist-pull, disturbance and handshake trials 10/10 on a real G1.

What they do

Whole-body motion trackers follow their reference closely, so they resist a person guiding the robot’s arm. ILA (Interface-Level Adapter) adds compliance without retraining the tracker, estimating wrenches or adding interaction sensors. It edits the reference the tracker receives, and the frozen tracker keeps handling balance and whole-body execution.

How it works

  • From the nominal reference and a history of proprioception and past tracker actions, ILA predicts a 32-step residual for the 14 arm joints. Trunk and leg references stay nominal.
  • Training pairs nominal motions (LAFAN1, AMASS, 100STYLE) with modified ones in which an arm moves toward a temporary pose, holds and returns. A simulated spring wrench pulls the hand accordingly.
  • A privileged teacher sees both references. A student that sees only the history is distilled from it and also predicts the interaction phase (none, yield, hold, recover) and per-arm intensity.
  • Fixed mappings of joint order, coordinates and timing plug the same adapter into other trackers.

Results

  • Simulation (243 matched guidance rollouts per method): target error drops 68.1% on MotionTracking, 63.6% on MimicLite and 75.4% on Sonic, with lower fall rates. No-force tracking error rises 0.5% on MotionTracking.
  • ILA gives stronger guidance and better recovery after release than finetuning each tracker. GentleHumanoid gets closer to the target (2.6 vs 5.7 cm) but moves the uninvolved arm more (10.6° vs 2.8°).
  • G1 hardware, 10 trials per task: the original trackers never reach the wrist-pull target and fail all disturbance and handshake trials. With ILA all three succeed 10/10, as do native MotionTracking compliance and GentleHumanoid.
  • Inference takes 0.51 ms on the G1’s CPU.

Limitations

Only the arms adapt. On Sonic, no-force tracking error grows more (5.36° to 5.85°), and how much nominal motion changes depends on the tracker. On hardware ILA matches, but does not beat, existing compliance methods.

Figures

From Figure 2: ILA adapts the reference while preserving the controller. (a) Pairs of nominal and modified references...
From Figure 2: ILA adapts the reference while preserving the controller. (a) Pairs of nominal and modified references supervise the residual correction. (b) A privileged teacher and a causal-history student share a decoder; at deployment the student adapts the motion intent and the frozen tracker performs whole-body execution and stabilization. (c) The same frozen adapter is reused with other frozen trackers through fixed observation and reference mappings.
From Figure 3: Guidance and motion preservation, lower is better on both axes. (a) Guidance target error versus no-fo...
From Figure 3: Guidance and motion preservation, lower is better on both axes. (a) Guidance target error versus no-force joint tracking error; hollow and filled markers show each frozen tracker without and with the same ILA. (b) Guidance target error versus whole-body recovery after release; the dashed line shows the non-dominated frontier.
From Figure 5: Real-world interaction tasks: a human-led handshake with the G1.
From Figure 5: Real-world interaction tasks: a human-led handshake with the G1.
Abstract

Humanoid motion tracking is fundamental to reproducing human motions on humanoid robots and serves as an important tool for teleoperation and motion data collection. However, directly tracking reference motions can be unsafe in constrained environments or during physical human--robot interaction, motivating the need for compliant motion tracking. We present ILA, an Interface-Level Adapter for compliant humanoid motion tracking without dedicated interaction sensors. ILA adapts motion references from the robot's observation history, allowing the robot to respond to physical guidance while preserving the capabilities of a pretrained motion tracker. As an interface-level module, ILA can be directly plugged into different pretrained motion trackers without retraining either the adapter or the tracking policies. ILA is trained entirely in simulation to predict reference corrections induced by physical interaction. We evaluate ILA with three humanoid motion trackers and achieve a 63.6--75.4% reduction in guidance-target error compared with their original policies. Matched simulation studies and hardware experiments further show that ILA improves responsiveness to human guidance while preserving nominal motion-tracking performance. These results demonstrate compliant guidance as a reusable, sensor-free extension to pretrained humanoid motion trackers.

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.