Humanoid Control  ·  Reinforcement Learning

LNP: Liquid Network Policy for
Humanoid Whole-Body Motion Tracking

  Anonymous Author(s) Affiliations withheld for double-blind review
Paper Code

Video

Muted for autoplay — unmute for the narrated audio.

Video overview Qualitative results of LNP across agile whole-body motion-tracking tasks — including aerial turns, forward leaps, run-and-kick, and fast walking. The clips show the policy tracking reference human motions in large-scale parallel simulation (IsaacLab) and deployed zero-shot on a real humanoid, illustrating both tracking accuracy and robustness to real-world observation latency.

Abstract

Tracking agile, whole-body human motion on a humanoid robot demands a control policy that can reason over both long temporal context and fast, instantaneous action–state dynamics, while remaining robust to the latency and timing jitter inherent to real hardware. We present LNP (Liquid Network Policy), a time-conditioned hybrid-memory recurrent policy for humanoid whole-body motion tracking. LNP combines a recurrent LSTM branch, which captures long-horizon temporal context, with a continuous-time liquid recurrent branch, which models rapid action–state dynamics. The two memory streams are fused through a time-aware gate that is explicitly parameterized by the control period, allowing the policy to adapt its effective memory horizon to the rate at which it is queried. We train LNP with a PPO-compatible truncated backpropagation-through-time procedure that carefully handles state recovery, overlapping windows, burn-in, and loss masking on valid segments. Across four challenging motion-tracking tasks, LNP attains higher success rates and lower tracking error than strong baselines, and it transfers to a real humanoid with substantially improved tolerance to observation latency.

LNP overview: architecture, training in simulation, real-world deployment, and performance.
Overview of LNP. A time-conditioned hybrid-memory recurrent policy couples a long-horizon LSTM branch with a fast liquid recurrent branch through a time-aware gate parameterized by the control period. The policy is trained in large-scale parallel simulation and deployed zero-shot on a real humanoid, achieving high tracking accuracy and strong robustness to observation latency.

Key Contributions

Hybrid Memory

A dual-branch recurrent policy that unifies long-horizon LSTM context with fast continuous-time liquid dynamics.

Time-Aware Gate

An interpolation gate parameterized by the control period adapts the effective memory horizon to the query rate.

Stable RL Training

A PPO-compatible TBPTT scheme with state recovery, overlapping windows, burn-in, and valid-segment loss masking.

Latency Robustness

Zero-shot deployment on a real humanoid with graceful degradation under 0–160 ms observation latency.

Method

LNP framework: human video input, feature extractor, actor with LSTM and liquid cell, and critic.
LNP framework. Reference motion is reconstructed from human video and provides the tracking target. At each control step, the actor encodes proprioception and the reference signal, routes them through the parallel LSTM and liquid recurrent branches, and fuses the two hidden states with the time-aware gate before the action head. The policy is optimized with a PPO-compatible TBPTT procedure.

The core of LNP is a learned interpolation between two complementary memory streams. Given the LSTM hidden state and the liquid-cell hidden state, a time-aware gate gt mixes them as a function of the control period Δτ:  gt = σ( at ⊙ Δτ + bt ).

When the policy is queried frequently (small Δτ), the gate favors the fast liquid branch to track instantaneous dynamics; when the effective control period grows—for instance, under observation latency—the gate shifts weight toward the long-horizon LSTM context. This explicit dependence on Δτ is what gives LNP its characteristic robustness to timing variation on real hardware.

Results in Simulation

We evaluate LNP on four agile motion-tracking tasks against strong recurrent and adversarial baselines. LNP consistently achieves the highest success rate and the lowest global tracking error (Eg-mpjpe), demonstrating the benefit of time-conditioned hybrid memory.

Task Method Success (%) ↑ Eg-mpjpe ↓
Aerial TurnLNP (Ours)100101.9
ASAP100147.5
AMP0170.9
MHC0180.1
Forward LeapLNP (Ours)100—
ASAP——
Run-and-KickLNP (Ours)100—
ASAP——
Fast WalkLNP (Ours)10088.8
ASAP——

Higher is better for success rate (↑); lower is better for tracking error (↓). Best results per task are highlighted. Metric definitions follow the paper: Eg-mpjpe denotes global mean per-joint position error.

Real-World Deployment

We deploy the trained policy zero-shot on a real humanoid. LNP transfers reliably to hardware, improving both success rate and Cartesian tracking error over the strongest simulation baseline.

Task Method Success (%) ↑ Tracking Error (mm) ↓
Jump-ForwardLNP (Ours)9059.7
ASAP7068.1
Fast WalkLNP (Ours)10050.1
ASAP10059.8

Real-hardware evaluation. LNP attains higher success and lower end-effector tracking error than ASAP, confirming that the time-aware hybrid memory transfers to the physical platform.

Simulation vs. Real-World

Side-by-side qualitative comparison of LNP in large-scale parallel simulation (IsaacLab) and deployed zero-shot on the real humanoid. For both the jump-forward and fast-walk tasks, the real-world rollout closely matches the reference motion produced in simulation, illustrating reliable sim-to-real transfer.

Jump-Forward

Simulation
Real-World

Fast Walk

Simulation
Real-World

Robustness to Observation Latency

A central advantage of the time-aware gate is graceful behavior under delayed observations. As injected observation latency increases from 0 ms to 160 ms, baselines degrade sharply, whereas LNP maintains high tracking quality by shifting memory weight toward its long-horizon branch. This makes LNP markedly more dependable for real-time control on hardware with non-negligible sensing and communication delays.

0–160 ms tested

Performance evaluated across the full latency sweep relevant to real deployment.

Graceful degradation

LNP retains tracking quality where adversarial and single-memory baselines collapse.

Timing adaptivity

The gate re-balances fast and slow memory as the effective control period changes.

Conclusion

We introduced LNP, a time-conditioned hybrid-memory recurrent policy for humanoid whole-body motion tracking. By fusing a long-horizon LSTM branch with a fast continuous-time liquid branch through a control-period-aware gate, and by training the policy with a carefully designed PPO-compatible TBPTT procedure, LNP achieves state-of-the-art tracking accuracy in simulation, transfers zero-shot to a real humanoid, and remains robust to observation latency. We believe time-conditioned memory is a promising direction for reliable real-time control of agile humanoid systems.

BibTeX

@inproceedings{lnp_anonymous,
  title     = {LNP: Liquid Network Policy for Humanoid Whole-Body Motion Tracking},
  author    = {Anonymous Author(s)},
  booktitle = {Proceedings of the Anonymous Conference (Under Review)},
  year      = {2026},
  note      = {Double-blind submission. Author and affiliation information withheld.}
}