Hyeonjun Cho

Research · 2026 · Manuscript in preparation

Recovery-Aware LPB for RMP-Induced OOD States

A policy bridge that detects states created by reactive safety interventions and returns the robot to task-relevant behavior.

Problem

Reactive safety layers such as RMPflow keep a manipulator away from obstacles, but the avoidance motion can leave the robot in states that never appeared in the demonstrations. A diffusion policy queried from such a state has no reliable reference to return to, so a single safe intervention can turn into a task failure. The question was whether a latent-guidance method, Latent Policy Barrier (LPB), could be made aware of these avoidance-induced states without retraining the base policy.

Key result

91.3%

Recovery-aware LPB. 146 successes in 160 trials of the proposed method across eight OOD start configurations.

Contribution

I designed the recovery bridge and evaluated the proposed method in 160 RB10 trials. It completed 146 trials (91.3%); the original LPB reached 64.4% under the same intervention setting.

My Role

  • Designed the recovery-state definition and recovery-aware policy bridge.
  • Built the RB10 and ROS 2 evaluation pipeline for intervention and recovery trials.
  • Analyzed failure modes across baseline, fine-tuning, original LPB, and the proposed method.

Team Contribution

  • The broader team developed and maintained the robot platform, task setup, and safety procedures.
  • Research mentors and I refined the experimental protocol and manuscript framing.

Representative experiment

Video 1. A short representative rollout from the real-robot intervention-and-recovery evaluation.

Method

I reimplemented Latent Policy Barrier and augmented its latent-dynamics model with post-avoidance recovery transitions collected from real RMP interventions, keeping the base diffusion policy fixed. The framework combines diffusion-policy references, RMPflow reactive avoidance triggered by proximity signals, and LPB latent guidance that steers the robot back toward the demonstration distribution. An RB10 / ROS 2 pipeline synchronizes policy references, RMP-modified executions, dual-camera observations, proximity signals, and LPB internal states for every trial.

Diagram showing a task policy, RMP safety intervention, OOD state, and recovery-aware policy bridge returning control to the task policy.
Figure 1. The RMP safety layer can move the robot outside the demonstration distribution. The recovery bridge returns control from those states to task execution.
  1. Proximity-triggered RMP intervention
  2. OOD state identification
  3. Recovery-aware policy bridge
  4. Real-robot rollout evaluation

Evaluation results

ConditionResultNote
Recovery-aware LPB91.3%146 successes in 160 trials of the proposed method across eight OOD start configurations.
Original LPB64.4%Evaluated under the same intervention setting.
Fine-tuning63.8%Direct adaptation baseline.
Base policy25.0%No recovery-specific adaptation.

Across eight real-robot OOD start configurations, the recovery-aware bridge succeeded in 146 of 160 trials (91.3%). Under the same intervention setting, the recorded comparison rates were 64.4% for original LPB, 63.8% for policy fine-tuning on the same recovery demonstrations, and 25.0% for the base policy. Baseline trial counts are not stated because their denominators are not confirmed in the available records.

Experiment evidence

Video 2. RMPflow deflects the RB10 around an introduced obstacle before LPB-guided task execution resumes. The 55-second excerpt masks every visible third-party face.
Video 3. Representative guided recovery trace with synchronized observations, trajectory views, and the LPB cost timeline.

Additional experiment evidence

Video 4. Short excerpt from the continuous RB10 LPB inference session, including workpiece repositioning between repeated rollouts.
Video 5. A second trial shows the recovery cost falling below the decision threshold as the robot returns toward task execution.
Video 6. Continuous real-robot LPB inference session across repeated rollouts (3 min 27 s at 1.7× speed). Hyeonjun Cho's face is shown with permission; one third-party face is masked.

Limitations

  • Results are specific to the evaluated task distribution and intervention protocol.
  • The manuscript and broader task-suite evaluation are still in preparation.

Materials

Stack. Python, PyTorch, ROS 2, RB10, RMP, Imitation Learning.

Code and materials. Public OOD tooling fork · Request full implementation access