Humanoid robots learn to yield: ILA cuts guidance-target error by 63.6%–75.4% across three trackers without retraining
Synopsis
When a person physically guides a humanoid robot that is tracking a prescribed motion, following the reference exactly conflicts with yielding to that guidance. ILA is an interface-level adapter that predicts corrections to the motion reference from the robot's observation history and inserts them ahead of a pretrained tracker, leaving the tracker's architecture and parameters unchanged and adding no dedicated interaction sensor. The same adapter reduces guidance-target error by 63.6%–75.4% across three independently trained trackers, and simulation plus G1 hardware experiments show better guidance response while largely preserving nominal motion tracking.
Interpretation
A pretrained humanoid motion tracker can accept physical human guidance while keeping its existing motion-execution ability, without being retrained. Earlier learned compliance approaches required retraining or redesigning the motion controller, while analytical retrofits relied on explicit wrench estimation and a prescribed force–motion relation. Adding ILA to the source tracker MT-Std reduces guidance-target error by 68.1%, alongside lower interaction work, root displacement and fall rate; no-force tracking error rises by only 0.5%.
ILA rewrites only the reference sent to the tracker, placing compliance at the reference layer rather than in the execution layer. The adapter predicts a multi-step residual over 14 arm joints and inserts the corrections through a mask, leaving trunk and leg references intact; zero correction restores the original reference. Training happens entirely in simulation with paired nominal and modified references that encode yielding, holding and recovery; the same adapter connects to MimicLite and Sonic through deterministic mappings of joint ordering, coordinate conventions and reference timing, with adapter and tracker parameters fixed.
Perspective
The demonstrated scope is matched guidance experiments with three trackers in simulation, plus ten physical trials each on G1 hardware for wrist pulling, unilateral disturbance and continuous handshaking; standalone adapter inference on the G1 controller takes 0.51 ms in a separate asynchronous trajectory-generation process, small relative to the 20 ms control cycle.
Cross-tracker reuse is validated only on MT-Std, MimicLite and Sonic; compared with full-policy finetuning, ILA shows a larger no-force tracking-error increase on Sonic together with substantially stronger guidance. Differences across motions are larger, with FightAndSports1 showing the highest fall rate and recovery error, while target error varies little with training seed, with a coefficient of variation of 4.8%.
