PhaDeR

Learning from Corrections and Interventions / 2026

PhaDeRPhase-Conditioned Failure Detection and Recovery
for Long-Horizon Humanoid Manipulation

Nga Teng Chan1,2 Pengxu Chang3,2 Yongquan Chen4,2 Xi Zhao2 Qifeng Chen1,* Dedong Li2,*

1 The Hong Kong University of Science and Technology

2 Shanghai Jushen Technology

3 Shanghai Jiaotong University

4 Zhejiang University

* Corresponding authors

Work in progress

Watch demonstrationsPaper overview
SIMULATION · UNITREE G11× SPEED

FROM FAILURE TO CONTINUATION

A missed grasp.
A second chance.

  1. 01
    Detect

    Recognize failure in the context of the task.

  2. 02
    Recover

    Apply a correction learned from human interventions.

  3. 03
    Resume

    Return to the original policy and complete the handover.

One policy. A dedicated recovery adapter.
Recovery in action. A natural missed grasp, a learned regrasp, and a completed Handover task. This continuous execution is shown at original speed; the first-person Handover demo below shows the same run.
See the missed grasp and successful regrasp Close-up replay · 0.5× speed · source 1.6–13.0 s
A continuous crop of the overview, slowed to 0.5×. The initial grasp loses contact; recovery retries until it secures and lifts the box.
  1. The hand has missed the box, which remains on the source table
    2.90 s · MISSED GRASP Contact is lost. The box stays.

    The initial grasp loses contact with the box.

  2. The robot holds the box clear of the source tabletop
    12.50 s · GRASP RECOVERED A new grip. A successful lift.

    Control returns to the nominal policy at 12.90 s.

  3. First-person view of the box reaching the basket at the end of the Handover task
    16.44 s · TASK COMPLETED The handover is complete.

    The nominal policy brings the box to the basket.

THE IDEA

Don’t just fix the grasp.
Make the task continuable.

Long-horizon humanoid manipulation rarely goes perfectly. A missed grasp can derail everything that follows. Recognizing the failure is only the first step: a correction must leave the robot in a state from which it can continue.

PhaDeR brings together phase-conditioned failure detection and a recovery-only adapter learned from human corrections. The nominal policy stays frozen. When a failure is detected, the adapter takes over; after readiness checks, control returns to the original policy.

01 THE SYSTEM IN ACTION

Detect. Recover. Resume.

Three tasks. The same corrective loop. Watch selected executions move from an acquisition failure to learned recovery and nominal task completion.

NATURAL MISSED GRASP · SAME RUN AS OVERVIEW

Retry the acquisition. Complete the handover.

Task completed
NOMINAL EXECUTIONEGOCENTRIC VIEW

The same execution as the overview, from the robot’s view. The initial grasp loses contact. The recovery adapter retries, secures and lifts the box, then the frozen nominal policy resumes and completes the handover.

Selected development successes in simulation, not an estimate of recovery success rate. Handover shows a natural missed grasp; the other two examples use controlled grasp perturbations. Task-specific deployment monitors and rise-based triggers differ from the shared-detector replay benchmark below. Handover success uses the original target–basket contact criterion; it does not require full hand release. View earlier development keyframes and trigger traces ↗

Earlier diagnostic comparisonGrasp-stage excerpts with and without recovery
Nominal policy only
The grasp fails and the box is left behind.
With PhaDeR recovery
The robot grasps the box again and resumes transport.
Left: 0–9 s · Right: 0–15 s · Original speed

Grasp-stage excerpts from independent runs in the same scene, illustrating a failed grasp without recovery and a regrasp with recovery.

02 THE METHOD

Know the phase.
Choose the correction.

The same observation can mean different things during approach, grasping, or placement. PhaDeR uses a predicted task-phase belief to put failure evidence in context.

Execution historyVision · state · past actions
Phase + failureCausal, predicted context
Recover & resumeAdapter → frozen policy
01 / MONITOR

Understand the execution.

A frozen encoder combines V-JEPA 2 visual features, proprioception, and lagged actions. A causal GRU estimates task phase from execution history.

02 / CORRECT

Learn from interventions.

A phase-conditioned detector recognizes failure. A rank-16 LoRA adapter, trained on human corrections, activates only during recovery.

03 / CONTINUE

Return control deliberately.

Simulator-assisted readiness checks gate the handoff. The original Ψ₀ policy [1] resumes with its parameters unchanged.

PhaDeR architecture: observation, robot state, and action history feed an encoder, task-phase estimator, and failure detector. A recovery adapter activates on failure and hands back to the frozen action expert. Third-person frames illustrate a missed grasp, retry, lift, handoff, and resumed transport.Open full figure ↗
The PhaDeR framework and an illustrative recovery sequence from the manuscript. The pictured sequence demonstrates resumed transport, rather than completed placement.
What stays frozen, and what is learned?

The execution encoder is pretrained on nominal demonstrations and then frozen. The task-phase estimator and failure detector learn from role-specific supervision; the detector receives predicted phase probabilities, rather than annotated phases.

Task-specific Ψ₀ nominal policies remain frozen. Recovery adapters learn with flow matching, 20% nominal replay, and a preservation loss. The selected task demonstrations use 35 corrective segments for Mobile Pick & Place, 17 for Handover, and 15 for P&P HugContainer.

The current handoff uses simulator-assisted contact, lift, stability, and safety checks. A learned continuation gate is future work.

03 EXPERIMENTAL FINDINGS

Task phase adds useful context.

Across three tasks simulated in SIMPLE [2], predicted phase improves event detection in a matched GRU comparison, with no observed false alarms on the clean test trajectories.

88.2%

Event detection rate± 0.7% across three training seeds

+3.1pp

Gain from predicted phaseOver the matched phase-free GRU

0/ 72

Clean episodes with an alarmObserved on the held-out test set

Failure-monitoring benchmark

Full-trajectory replay
Event detection, false alarms, and conditional delays. Plus or minus denotes standard deviation across three independent GRU seeds.
Method Phase input Detection ↑ False alarms ↓ Median delay ↓
GRU detector None 85.1 ± 0.7% 0 / 72 22.3 ± 2.9
GT-phase GRU Annotated 87.2 ± 0.7% 0 / 72 21.7 ± 1.7
PhaDeR Ours Predicted 88.2 ± 0.7% 0 / 72 23.3 ± 2.4

126 test episodes · 65 failure events · 72 clean episodes. Mean ± SD over three GRU seeds. Delay is in control steps among detected events. Replay includes post-intervention observations, so these are event-recognition results, not a measurement of autonomous pre-intervention detection.

Detection rate is one part of the story. PhaDeR’s conditional delays are higher than the phase-free GRU’s; the detected event sets differ, so this is not a matched-event timing comparison.

Failure-score traces and egocentric keyframes for failed and successful executions across Handover, Mobile Pick and Place, and P and P HugContainerOpen full figure ↗
Selected failure-score traces and egocentric observations. These qualitative examples use task-specific monitors, unlike the shared-detector benchmark.

BEYOND LOCAL RECOVERY

A successful handoff is not the finish line.

Restored-state trials separate passing the local readiness gate from completing the original task. In Mobile Pick & Place, one additional local recovery did not translate into another successful continuation.

Six restored-state reports or trials per task; results are preliminary development outcomes.
Task Local recovery Continuation after handoff Full completion
Handover 2 / 6 2 / 2 2 / 6
Mobile Pick & Place 3 / 6 2 / 3 2 / 6
P&P HugContainer 2 / 6 2 / 2 2 / 6

Selected development outcomes; six reports or trials per task. Conditional continuation counts only trials that reached handoff.

04 AN ONGOING PROJECT

Toward corrections
that lead to completion.

PhaDeR connects failure recognition to task resumption. The next step is to make that connection more reliable: expand corrective data, learn handoff criteria from nominal continuation outcomes, and transfer to real humanoids.

More corrective experienceContinuation-aware handoffReal-world transfer

The current study is preliminary and simulation-based. Selected successes establish integration; they do not establish aggregate closed-loop reliability or superiority of probability-rise triggering.

References
  1. Ψ₀: An Open Foundation Model Towards Universal Humanoid Loco-Manipulation
  2. SIMPLE: Simulation-Based Policy Learning and Evaluation for Humanoid Loco-manipulation
Cite this work

Nga Teng Chan, Pengxu Chang, Yongquan Chen, Xi Zhao, Qifeng Chen, and Dedong Li. “PhaDeR: Phase-Conditioned Failure Detection and Recovery for Long-Horizon Humanoid Manipulation.” 2026.

BibTeX

@misc{phader2026,
  author = {Chan, Nga Teng and Chang, Pengxu and Chen, Yongquan and
            Zhao, Xi and Chen, Qifeng and Li, Dedong},
  title = {{PhaDeR}: Phase-Conditioned Failure Detection and
           Recovery for Long-Horizon Humanoid Manipulation},
  year = {2026},
  url = {https://tengisa1118.github.io/PhaDeR/}
}