UniReflex: Plug-and-Play Force Control for Pre-trained Generative Policy via Fast-Slow Reflex Coupling

Generative imitation learning policies excel at trajectory planning but lack closed-loop force regulation, while existing force-aware adaptations typically require costly retraining. We present UniReflex, a plug-and-play framework that endows frozen generative policies with active force control without backbone fine-tuning. By non-invasively intercepting deep latent representations from the action head, UniReflex drives a fast reflex network that decouples active force exertion from external interaction response. Inspired by human muscle co-contraction, this scheme predicts anisotropic stiffness for genuine active force exertion. Furthermore, UniReflex integrates an adaptive gating mechanism that enables seamless transitions between position-dominant planning and force-dominant execution. Real-world bimanual experiments demonstrate that UniReflex significantly improves contact stability and success rates while preserving original position accuracy. Notably, our approach achieves 25–66× lower per-step backward latency relative to joint training strategies on the evaluated backbones.

Generative imitation policies are typically position-based planners without closed-loop wrench regulation. UniReflex introduces a separately trained fast reflex that reads internal trajectory representations from the frozen slow policy and closes the loop on force at a high rate.

This section places UniReflex in the broader landscape of force-control methods. Existing approaches either add force as another policy output, rely on task-specific hybrid controllers, or train fast auxiliary modules together with the main backbone. UniReflex follows a different route: it keeps the pre-trained generative policy untouched and adds a lightweight reflex layer that translates its latent intent into contact-aware force regulation.

UniReflex uniquely enables active, variable-stiffness control while remaining plug-and-play with frozen generative policies via fast-only training.

UniReflex Architecture. We hook the deep feature map of the generative action head to extract an intent tensor. The reflex GRU consumes proprioception and this intent tensor to emit reference poses, reference wrenches, and stiffness encodings for decoupled force control. The system operates in a tri-rate closed loop: slow visual planning, fast reflex prediction, and PID force tracking.

The key idea is to separate what the robot should do from how contact should be regulated. The frozen policy continues to provide semantic motion intent from vision, while the reflex module reacts to high-frequency proprioception and contact signals. During free-space motion, the controller prioritizes pose tracking; after contact is detected, the gate shifts authority toward force and stiffness regulation.

The system diagram shows the full fast-slow coupling: latent features from the slow generative policy condition a reflex network, which predicts reference motion, desired wrench, and anisotropic stiffness for the low-level controller. This lets UniReflex respond to contact much faster than the visual policy can re-plan.

We evaluate five contact-rich bimanual benchmarks with frozen generative backbones (DP, \(\pi_0\), DreamZero) and force-aware baselines (RDP, ForceVLA, TA-VLA). Only UniReflex is trained; the slow policy stays frozen.

Following the paper, results are organized around six questions:

• Q1 Phase I / Phase II task performance with frozen backbones
• Q2 Comparison against monolithic force-aware baselines
• Q3 Component ablations (stiffness, wrench, gating)
• Q4 Recovery under dynamic disturbance and TCP bias
• Q5 Interaction-time force tracking
• Q6 Training throughput vs. joint fine-tuning

Five tasks cover sustained wiping, peeling, insertion, and impact/flip. Phase I is approach/alignment; Phase II is force-regulated contact.

Q1–Q2 Success rates on contact-rich benchmarks

Q1. Phase I stays within ~10 points of each frozen backbone; Phase II rises by 20–60 points on sustained-contact tasks. Q2. Against RDP / ForceVLA / TA-VLA, UniReflex keeps planner-level Phase I quality and matches or exceeds the best monolithic Phase II numbers.

Q3–Q4 Ablations and recovery

Q3. Stiffness / wrench / gating ablations under frozen DP.

Q4. Dynamic recovery 80–100% (0.87–1.28 s); TCP-bias Phase II 70–90%.

Q4–Q5 High-frequency reflex and force tracking

Q4–Q5. Fast corrections under disturbance; peel-force tracking closer than frozen baselines; measured \(F_x,F_y,F_z\) aligned with UniReflex references.

Q6 Training throughput

Q6. Training only the fast GRU gives about 25–66× lower per-step backward latency than joint fine-tuning, while updating only 0.188–0.375% as many parameters as the backbone.

We present 5 evaluation tasks. Each row corresponds to one task, featuring 3 columns: (1) Frozen backbones, (2) Force-aware baselines, and (3) UniReflex with different backbones. Use the left/right arrows or the dots below to switch between the 3 videos in each column.

These videos are meant to make the table results concrete. Frozen backbones often generate plausible motion but fail once contact must be maintained. Force-aware baselines can react to force, but may lose position precision or require tighter coupling with the policy. UniReflex aims to keep the backbone's motion plan while adding the fast corrections needed for stable physical interaction.

Comparison of force-centric execution. Use the arrows to switch between frozen backbones, force-aware baselines, and UniReflex.

This comparison focuses on the moment of contact rather than the full task rollout. The important difference is not only whether the robot reaches the target, but whether it can apply force smoothly without bouncing, stalling, or drifting away from the intended motion.

Q5. UniReflex predicts a task-dependent reference force instead of relying on a fixed contact heuristic. The tracking clip shows how the measured wrench follows this reference during execution, which is the mechanism behind the observed contact stability in the manipulation tasks.

Q4. The fast reflex runs at a higher rate than visual policy inference, so it can respond to disturbances while the slow planner continues to provide the global trajectory. This clip emphasizes the temporal advantage of the fast-slow design: corrections happen at the contact timescale, not only at the next camera-driven action update.