🤖 machine learning

RAPT: Model-Predictive Out-of-Distribution Detection and Failure Diagnosis for Sim-to-Real Humanoid Deployment

This paper introduces RAPT, a lightweight, self-supervised monitor that enables high-frequency out-of-distribution detection and interpretable, zero-shot failure diagnosis for humanoid robots during Sim-to-Real deployment, significantly reducing silent failures and hardware damage risks.

Original authors: Humphrey Munn, Brendan Tidd, Peter Bohm, Marcus Gallagher, David Howard

Published 2026-07-22

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Humphrey Munn, Brendan Tidd, Peter Bohm, Marcus Gallagher, David Howard

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are teaching a robot to walk by letting it practice millions of times inside a perfect, video-game world. In that digital playground, the robot learns to balance, run, and dance without ever tripping. But when you finally unplug it and let it step into the real world, things get messy. The real world has slippery floors, unexpected bumps, and sensors that sometimes lie. This is the tricky "Sim-to-Real" problem: a robot might act like a champion in the game but stumble immediately in reality because it hasn't seen these real-world glitches before. If the robot doesn't realize it's confused, it might keep trying to walk on a broken leg or push against a wall that isn't there, potentially crashing into itself or breaking its expensive parts. Scientists are trying to build a "safety net" that watches the robot, spots when it's entering a weird, dangerous situation it wasn't trained for, and hits the brakes before disaster strikes.

This is exactly what the paper "RAPT" tackles. The authors built a lightweight, super-fast watchdog system called RAPT (Recurrent Anomaly Probabilistic Trajectory Model) that runs alongside a robot's brain. Think of RAPT as a highly experienced coach who has watched the robot practice in the video game so much that they know exactly how it should move. When the robot starts moving in the real world, RAPT watches every joint and sensor in real-time. If the robot suddenly starts wobbling in a way that doesn't match its "perfect practice" memory—like if it gets pushed hard or steps on a slippery patch—RAPT instantly shouts, "Hey, that's not normal!" and triggers a safe stop.

What makes RAPT special is that it doesn't just say "stop"; it acts like a detective. It can pinpoint which part of the robot is acting weird and when it started acting weird. Even cooler, it can talk to a smart language assistant (an AI) to guess why the robot failed, like "Oh, the robot probably slipped on a wet floor" or "Someone pushed it," without needing to have seen that specific accident before. In their tests, the team trained RAPT in a massive simulation and then tried it on a real robot. The results showed that RAPT was much better at catching these dangerous moments than other methods, catching 37% more problems in the simulation and 89% of the problems on the real robot, all while rarely crying wolf when everything was actually fine. It's a step toward making robots that can safely explore the messy real world without breaking themselves.

Technical Summary: RAPT – Model-Predictive Out-of-Distribution Detection and Failure Diagnosis for Sim-to-Real Humanoid Deployment

Problem Statement

Deploying learned control policies for humanoid robots from simulation to reality (Sim-to-Real) carries significant risk. Policies that appear robust in simulation may confidently enter out-of-distribution (OOD) states upon deployment due to unmodeled dynamics, sensor noise, latency, payload changes, terrain variations, or implementation errors. These OOD events can lead to "silent failures," falls, self-collisions, or hardware damage.

Existing anomaly detection methods often fail to meet the specific constraints of high-rate humanoid control (50 Hz):

Computational Cost: Uncertainty and ensemble methods are often too expensive for real-time control loops.
Calibration Issues: Generative models (e.g., LSTM-VAEs) can suffer from likelihood miscalibration and input-complexity bias.
Lack of Interpretability: Many approaches provide only a binary stop signal, offering no insight into when, where, or why the execution deviated from nominal behavior.
State Coupling: Generic time-series methods often treat proprioceptive channels as independent signals rather than coupled robot states.

Methodology

The paper introduces RAPT (Recurrent Anomaly Probabilistic Trajectory Model), a lightweight, self-supervised monitoring system designed to run alongside a pre-trained control policy. RAPT operates in four stages:

1. Nominal Data Collection

RAPT collects nominal proprioceptive trajectories from expert policies executed in large-scale simulation (NVIDIA Isaac Lab). These trajectories define the "nominal behavior manifold" that the system must preserve during deployment. Observations are normalized prior to training.

2. Probabilistic Recurrent Trajectory Model

RAPT is trained as a self-supervised recurrent reconstruction model. The architecture includes:

Residual State Encoder: Processes observation history.
GRU-based Temporal Bridge: Captures temporal dependencies.
Latent Bottleneck: Compresses the state representation.
Probabilistic Decoder: Outputs a diagonal Gaussian reconstruction distribution $[\mu_t, \log \sigma^2_t]$ .

The model minimizes the Negative Log-Likelihood (NLL) of nominal observations. This formulation captures heteroscedastic aleatoric uncertainty, allowing dimensions with naturally higher variability to have larger predictive variance, while penalizing deviations in consistent dimensions more heavily. For velocity tracking tasks, a forward-dynamics variant (conditioning on $o_{t-1}, a_{t-1}$ to predict $o_t$ ) is used; otherwise, pure reconstruction is preferred.

3. Simulation-Trained, Deployment-Calibrated OOD Detection

At deployment, RAPT evaluates per-dimension predictive deviations. To handle the Sim-to-Real gap, RAPT supports two calibration modes:

Simulation Calibration: Measures deviations from the simulation-trained distribution, making residual Sim-to-Real mismatches visible.
Real-World Calibration: Uses a brief verified nominal run to adapt to static offsets (sensor noise, latency, hardware dynamics), enabling detection of new OOD events relative to the specific hardware.

Detection utilizes a hierarchical gating mechanism:

Local Gate: Detects unusually large per-dimension deviations ( $\ell_{t,i} > \tau_{max} + 5\sigma_i$ ).
Global Gate: Detects broad shifts in mean predictive deviation ( $\bar{\ell}_t > \tau_{global} + 3\sigma_{global}$ ).
Hybrid Safety Envelope: Optionally paired with a deterministic range detector. The final anomaly decision is the logical OR of local, global, and range gates.

4. Failure Localization and Diagnosis

Temporal Saliency: Upon anomaly detection, RAPT computes gradients of the final NLL with respect to a history window using integrated gradients through the recurrent model. This produces a spatio-temporal attribution map, highlighting when and which proprioceptive channels first departed from nominal behavior.
Semantic Diagnosis: A multi-modal Large Language Model (LLM) receives structured evidence (temporal saliency summaries, joint-kinematic trajectories, task context, and visual keyframes) to perform zero-shot classification of failure causes (e.g., collisions, sensor faults, terrain shifts) without requiring labeled failure data during detector training.

Key Contributions

Simulation-Trained, Deployment-Calibrated Monitoring: RAPT learns from large-scale simulation and calibrates thresholds using either simulation or brief real-world data, supporting both debugging and hardware-time OOD detection.
Low-FPR OOD Detection for High-DOF Humanoids: By using probabilistic predictive-deviation modeling with calibrated per-dimension and global gates, RAPT outperforms high-frequency-compatible baselines under strict false-positive constraints.
Localized Evidence for Failure Diagnosis: The system generates per-dimension NLL signals and temporal saliency, enabling post-hoc semantic diagnosis of failure modes without labeled failure data.

Experimental Results

The authors validated RAPT on a Unitree G1 humanoid in both NVIDIA Isaac Lab and on physical hardware.

Simulation Performance: Across four tasks (velocity tracking, motion mimicry, ballistic throwing) and 15 OOD categories, RAPT improved the True Positive Rate (TPR) by 37% over the strongest baseline at a 0.5% episode-level False Positive Rate (FPR).
Hardware Performance: On a physical Unitree G1 across 78 trials, RAPT achieved an 89% TPR with fewer false positives than high-frequency-compatible baselines.
Diagnosis Accuracy: On a challenging OOD subset, RAPT achieved 75% semantic failure diagnosis accuracy across 21 failure categories using the zero-shot LLM approach.
Latency: The monitoring pipeline operates at approximately 1.6ms, compatible with 50 Hz control loops.

Significance and Claims

The paper positions RAPT not as a replacement for domain randomization or system identification, but as a complementary layer that monitors the residual mismatch and unforeseen OOD events that persist after transfer.

Safety & Debugging: RAPT provides an observable deployment-time signal for safety monitoring and debugging without modifying the underlying controller. It triggers predefined safety responses (e.g., safe stop, controlled fall, recovery) upon OOD detection.
Interpretability: By moving beyond binary stop signals to localized, per-dimension evidence and semantic reasoning, RAPT addresses the "black box" nature of failure in learned policies, allowing operators to distinguish between external disturbances, actuator degradation, and sensor faults.
Reproducibility: The authors commit to releasing the training pipeline, labeled real-world datasets, and C++ deployment code to support future benchmarking.

The work demonstrates that a lightweight, self-supervised probabilistic model can effectively bridge the gap between simulation and reality, offering both immediate safety interventions and post-hoc diagnostic insights for complex humanoid systems.