🤖 Anonymous Submission  Â·  Under Review

UniIntervene

Agentic Intervention for Efficient Real-World Reinforcement Learning

An agentic intervention model that detects unproductive exploration and autonomously recovers the policy toward high-value states — turning human intervention into a value-aware recovery process for real-world robot RL.

Anonymous Author(s)
Affiliations hidden for double-blind review
UniIntervene overview teaser
UniIntervene replaces human intervention in HiL-RL with a value-aware critic that detects unproductive rollouts and a recovery policy that restores progress — yielding higher success with fewer human interventions on real-world manipulation.
+8.6%
Avg. success rate
vs. SOTA HiL-RL
−57%
Human interventions
vs. HiL-SERL
88%
Average task
success rate
14.6%
Intervention rate
(34.3% for HiL-SERL)
5
Real-world
manipulation tasks

01 Abstract

Human-in-the-loop reinforcement learning (HiL-RL) has emerged as an effective paradigm for real-world robotic manipulation, enabling online policy improvement with human guidance. However, current HiL-RL frameworks remain intervention-intensive, relying on frequent human corrections to redirect the policy out of unproductive exploration, which incurs high labor cost and limits real-world scalability.

To address this, we propose UniIntervene, an agentic intervention model that detects unproductive exploration and autonomously recovers the policy toward high-value states, taking over the bulk of interventions from human operators. UniIntervene first performs future-conditioned action-value estimation, predicting the latent consequence of the current action and evaluating its induced value, which provides a more stable progress signal. Building on this, a temporal value-risk critic aggregates recent value dynamics and triggers intervention when the estimated value exhibits sustained stagnation or degradation. When intervention is required, UniIntervene retrieves a high-value recovery target from a memory of past intervention episodes and produces executable corrective actions through a goal-conditioned recovery policy.

In this way, UniIntervene turns intervention from passive human correction into a value-aware recovery process for efficient real-world RL. Extensive experiments on diverse real-world manipulation tasks demonstrate that UniIntervene improves the average success rate by 8.6% while reducing human interventions by 57% relative to state-of-the-art HiL-RL baselines.

02 Overview Video

A walkthrough of UniIntervene detecting value stagnation and autonomously recovering on real-world tasks.

03 How UniIntervene Works

UniIntervene internalizes the intervention decision: it continuously asks whether the ongoing rollout is making productive progress, and autonomously redirects the policy when it is not.

UniIntervene method pipeline
Overview of UniIntervene. (a) A Qwen-VL backbone feeds a Latent Future Head with twin-critic and temporal value-risk supervision, and a Recovery Action Head producing corrective actions arec. (b) A memory buffer pairs past intervention states with retrieved high-value goals. (c) The temporal value-risk Rt triggers intervention upon sustained stagnation. (d) UniIntervene plugs into the real-world RL loop, overriding at with arec when triggered and contributing recovery transitions to the replay buffer.
1

Future-conditioned Action-Value

A model-based critic predicts the latent consequence of the current action and evaluates the value it induces — a far more stable progress signal under sparse rewards than scoring a single observation. Supervised by a frozen V-JEPA2 encoder and a progress-aligned proxy value function.

2

Temporal Value-Risk Trigger

Rather than firing on any low value, a risk critic aggregates recent value dynamics over a sliding window. It triggers only on sustained stagnation or degradation, filtering out the transient dips that naturally occur during contact, alignment, and regrasping.

3

Memory-guided Recovery

When triggered, UniIntervene retrieves a high-value recovery target from a memory of verified past recoveries, then a goal-conditioned policy decodes corrective actions toward it. Memory supplies where to recover; the policy learns how to get there — no human takeover required.

04 Real-World Task Suite

Five tasks on a UR7e arm spanning multi-object interaction, contact-rich assembly, and non-rigid manipulation — demanding pose robustness, precise contact, long-horizon correction, and recovery from low-value states.

The five real-world manipulation tasks
The five real-world manipulation tasks, shown from the third-person workspace camera during execution.
Multi-object

Pick Eggplant

Grasp the eggplant among distractors and lift it clear, leaving distractors undisturbed.

Contact-rich

Tube Insertion

Align and seat a flexible tube into a fixed port with sub-centimeter precision.

Contact-rich

RAM Insertion

Align a memory module to a slot and press it home — failures are visually subtle.

Contact-rich

Wipe Whiteboard

Wipe a marked region while maintaining contact force on a vertical surface.

Non-rigid

Fold Towel

Fold a flat towel along a target line through deformable, low-value regrasping states.

05 UniIntervene in Action

Qualitative rollouts of the full system: UniIntervene detects stagnation, takes over from the policy, and recovers toward task completion — autonomously.

Contact-rich Tube Insertion
Contact-rich RAM Insertion
Contact-rich Wipe Whiteboard

06 Head-to-Head Comparisons

Same task, same platform. UniIntervene (ours) recovers from low-value states and completes the task, while the HiL-SERL baseline stalls or requires repeated human correction.

UniIntervene (Ours)
HiL-SERL (Baseline)

07 Results

UniIntervene attains the best success rate on every task and the lowest overall intervention rate — less than half of HiL-SERL.

Method Pick Eggplant Contact-rich Fold Towel Average
SR / IR Tube Ins. RAM Ins. Wipe SR / IR
π0.5 (SFT) 95 / – 30 / – 10 / – 65 / – 70 / – 54 / –
HiL-SERL 90 / 28.7 60 / 30.2 85 / 32.3 85 / 30.5 85 / 49.8 81 / 34.3
HiL-SERL + FA-RL 85 / 20.4 60 / 22.1 75 / 27.9 80 / 21.9 85 / 30.9 77 / 24.6
HiL-SERL + UniIntervene 95 / 10.0 70 / 15.8 95 / 12.1 90 / 10.9 90 / 24.1 88 / 14.6

SR: success rate (%, higher is better). IR: human intervention rate (%, lower is better). Best results in bold.

Best on every task

UniIntervene tops success rate across all five tasks, with the largest gains on contact-rich settings where prior triggers struggle.

Value trend > failure events

Triggering on the temporal trend of estimated value detects stagnation more sensitively than failure prediction, especially when failures are visually subtle.

Half the human cost

Intervention rate drops to 14.6% — less than half of HiL-SERL (34.3%) and well below FA-RL (24.6%).

Ablation Study

Key components ablated on Pick Eggplant and RAM Insertion.

Variant Q-Loss ↓ Int. F1 ↑ SR (%) ↑ IR (%) ↓
w/o Future Prediction0.0050.8789015.8
w/o Value Prediction0.8458518.7
w/o Temporal Value-Risk0.0040.8328516.9
w/o Memory Goal0.0040.8828516.1
UniIntervene (full)0.0040.8829511.1

08 Value-Aware Recovery

Across episodes, UniIntervene follows a consistent stagnation → trigger → recovery pattern: it waits for sustained stagnation, fires at the value minimum, then climbs smoothly to success.

Case study of value-aware recovery
UniIntervene enables value-aware recovery in real-world tasks. On RAM Insertion and Fold Towel, the model detects sustained low-value stagnation, triggers corrective recovery at the value minimum, and restores task progress toward successful completion.

09 Hardware Setup

Real-world hardware setup
A UR7e arm with a parallel-jaw gripper, a wrist camera and a fixed third-person camera, and a SpaceMouse for corrective takeover. The same platform and intervention interface are shared by UniIntervene and all baselines.

10 BibTeX

@unpublished{uniintervene2026,
  title  = {UniIntervene: Agentic Intervention for Efficient
            Real-World Reinforcement Learning},
  author = {Anonymous Author(s)},
  year   = {2026},
  note   = {Under review}
}