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.
Figures & tables
Fig. 2: ILA adapts the reference while preserving the controller. (a) Nominal–modified reference pairs supervise the residual Δr∗=m⊙(rmod−rnom) . (b) O denotes the observation history and nominal reference (ht,rt) . A privileged teacher and a causal-history student share a decoder; at deployment, the student adapts motion intent and the frozen tracker performs whole-body execution and stabilization. (c) The same frozen adapter is reused with other compatible frozen trackers through deterministic observation and reference mappings.
Method
Target ↓
Work ↓
Inactive ↓
Root ↓
Recovery ↓
Fall ↓
(cm)
(J)
( ∘ )
(m)
( ∘ )
(%)
GentleHumanoid [ 5 ]
2.569
7.461
10.580
1.420
6.611
2.88
MT-Comp [ 27 ]
7.079
9.677
3.206
0.993
0.641
0.00
ILA (Ours)
5.684
10.859
2.815
1.246
0.619
0.00
MCC [ 8 ]
11.214
13.435
3.028
1.573
1.335
4.53
CHIP-R [ 7 ]
9.851
13.384
6.750
1.074
0.587
0.00
TABLE I: Matched SR-AOW v1 comparison on G1 over identical Dance1+Walk1 conditions. CHIP-R denotes our reimplementation.
Fig. 3: Guidance and motion preservation; lower is better on both axes. (a) Guidance target error versus no-force joint tracking RMSE against each run’s original reference. Hollow/filled markers denote each frozen tracker without/with the same ILA; native startup schedules differ across trackers. (b) Guidance target error versus whole-body post-release recovery under matched SR-AOW v1 conditions on G1; the dashed line shows the non-dominated frontier.
Method
Target
Work
Inactive
Root XY
Recovery
Fall
No-force tracking
(cm)
(J)
( ∘ )
(m)
( ∘ )
(%)
( ∘ )
MT-Std [ 27 ]
17.801
13.675
2.749
1.642
1.122
2.88
4.379
MT-FT
12.780
12.004
2.836
1.358
0.808
2.47
4.624
MT-Std+ILA (Ours)
5.684
10.859
2.815
1.246
0.619
0.00
4.400
MimicLite [ 41 ]
10.182
9.495
3.276
1.440
2.409
5.76
3.519
MimicLite-FT
9.856
9.251
3.349
1.493
1.931
5.35
3.724
TABLE II: Compliance retrofit and cross-tracker reuse. FT finetunes the full target tracking policy. Every +ILA row uses the same frozen MT-Std-trained adapter with the tracker fixed. Guidance uses matched SR-AOW v1 conditions; no-force tracking uses matched settings within each tracker. MT-Std and MT-Std+ILA guidance values follow Table I . Bold denotes the minimum within each tracker group.
Fig. 4: Cross-tracker reference adaptation under the same Walk1-left-20 N guidance condition. The thin orange wrist-to-sphere line indicates the applied external guidance force toward the orange target. Magenta denotes the nominal reference. With ILA, interaction-conditioned corrections shift the reference to the cyan pose, allowing the frozen tracker to yield toward the target. Columns compare each tracker without and with the same MT-Std-trained ILA; both adapter and tracker parameters remain fixed.
Fig. 5: Real-world interaction tasks. (a) Force-gauged wrist pull to a prescribed target. (b) Unilateral disturbance while the other hand maintains object support. (c) Human-led handshake. Table III reports task metrics.
Method
Pull force ↓ (N)
Disturbance success ↑
Handshake success ↑
MT-Std
NR @74.48
0/10
0/10
MT-Std + ILA (Ours)
16.88
10/10
10/10
MimicLite
NR @69.64
0/10
0/10
MimicLite + ILA (Ours)
10.72
10/10
10/10
Sonic
NR @77.42
0/10
0/10
Sonic + ILA (Ours)
18.64
10/10
10/10
TABLE III: Real-world interaction results over ten physical trials per method and task. Methods and sources follow Tables I and II .
Source
Target (cm)
Work (J)
Inactive ( ∘ )
Root (m)
Fall (%)
Direct
14.83 ± 0.08
16.13 ± 0.18
3.03 ± 0.02
2.08 ± 0.03
7.4 ± 0.1
IK
11.86 ± 0.12
14.70 ± 0.14
3.18 ± 0.13
1.76 ± 0.05
5.7 ± 0.2
Paired-reference (Ours)
7.99 ± 0.32
12.68 ± 0.62
2.68 ± 0.05
1.50 ± 0.10
3.8 ± 0.4
TABLE IV: Residual-supervision sources on MT-Std checkpoint 0509 (separate from the main benchmark). Values are mean ± standard deviation across three seeds and five motions.
Fig. 6: Controlled interventions on the latent z . (a) Latent substitutions assess the role of interaction context in guidance. (b) Phase-dependent residual activation. (c) Side-dependent adaptation. (d) Intensity-dependent residual magnitude. Reported intervals are 95% cluster bootstrap CIs (10,000 resamples; (a) n=243 target clusters across evaluation seeds 101–103, aggregated over three force levels; (b) n=48 clusters, (c) n=117 mirror pairs, (d) n=243 conditions); the shaded band in (d) is the 10–90% range across all evaluated interaction conditions.