💻 computer science

FAVLA: A Force-Adaptive Fast-Slow VLA model for Contact-Rich Robotic Manipulation

FAVLA is a force-adaptive Vision-Language-Action model that decouples slow perception planning from fast, variable-frequency contact-aware control to overcome the limitations of unified-frequency fusion and significantly improve reactivity and success rates in contact-rich robotic manipulation.

Original authors: Yao Li, Peiyuan Tang, Wuyang Zhang, Chengyang Zhu, Yifan Duan, Weikai Shi, Xiaodong Zhang, Zijiang Yang, Jianmin Ji, Yanyong Zhang

Published 2026-03-02

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Yao Li, Peiyuan Tang, Wuyang Zhang, Chengyang Zhu, Yifan Duan, Weikai Shi, Xiaodong Zhang, Zijiang Yang, Jianmin Ji, Yanyong Zhang

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 trying to teach a robot to perform a delicate task, like plugging a USB drive into a computer or screwing a tiny gear into place. These tasks are "contact-rich," meaning the robot has to actually touch things, feel resistance, and adjust instantly if something gets stuck.

The paper introduces a new robot brain called FAVLA. To understand why it's special, let's look at how robots used to think versus how FAVLA thinks.

The Old Way: The "Slow-Motion" Robot

Imagine a robot that takes a photo, thinks about what to do, and then moves. But here's the catch: it thinks and moves at the same slow speed.

The Problem: Cameras are like slow-motion cameras (taking 15 pictures a second), but force sensors (which feel pressure) are like high-speed cameras (taking 1,000 readings a second).
The Result: To make them work together, the old robots had to slow down the force sensors to match the camera. It's like trying to listen to a fast drum solo by only hearing every third beat. By the time the robot realizes it's pushing too hard and might break the gear, it's already too late. It's "open-loop," meaning it's guessing its way through the contact phase without real-time feedback.

The FAVLA Solution: The "Brain and Reflex" Team

FAVLA solves this by splitting the robot's mind into two distinct parts that work at different speeds, just like a human does.

1. The Slow Brain (The VLM)

Think of this as the Strategist.

What it does: It looks at the big picture. It reads the instructions ("Plug in the USB"), looks at the scene, and plans the general path.
Speed: It works slowly and deliberately. It doesn't need to react to every tiny vibration; it just needs to know the goal.
Superpower: It can predict the future. It looks at the current situation and guesses, "Hey, in a split second, we're going to hit a hard surface. Get ready!"

2. The Fast Reflex (The Action Expert)

Think of this as the Reflex.

What it does: It handles the actual movement and the "feel." It's the part that jerks your hand back if you touch a hot stove.
Speed: It works incredibly fast, reacting to the force sensors thousands of times a second.
Superpower: It doesn't just wait for orders. It listens to the "Strategist" for the plan, but then it uses its own high-speed "ears" (force sensors) to make micro-adjustments instantly.

The Secret Sauce: The "Force Adapter" and "Smart Scheduling"

FAVLA has two clever tricks that make this team work perfectly:

The Force Adapter (The Direct Line): In older robots, force data was just another piece of text added to the robot's "thoughts." In FAVLA, the force data is injected directly into the Reflex's muscles. It's like giving the reflex a direct phone line to the nerves, bypassing the brain's slow processing. This allows the robot to correct a mistake before it becomes a disaster.
The Smart Scheduler (The Traffic Cop): This is the most creative part. The "Strategist" predicts when the robot is about to hit something.
- Flying through the air? The Reflex relaxes and moves slowly (saving energy).
- About to touch the gear? The Strategist yells, "Contact imminent!" and the Reflex instantly speeds up to maximum sensitivity.
- Why this matters: It's like a driver who cruises at 60 mph on an empty highway but instantly switches to 10 mph and hyper-focuses when entering a crowded parking lot. The robot only goes "fast mode" when it actually needs to.

The Real-World Result

The researchers tested this on real robots doing tricky jobs like assembling gears and wiping boards.

Success Rate: The new robot succeeded 80.8% of the time, while the best previous robots only got about 67%.
Safety: The new robot was much gentler. It applied less force, meaning it was less likely to crush the delicate parts it was trying to assemble.

In a Nutshell

FAVLA is like giving a robot a human-like nervous system. It has a slow, thoughtful brain for planning and a fast, reactive body for feeling and adjusting. By letting the "reflexes" run at high speed only when necessary, the robot can handle delicate, touch-heavy tasks with a level of precision and safety that previous robots simply couldn't achieve.

1. Problem Statement

Current Vision-Language-Action (VLA) models struggle with contact-rich manipulation tasks (e.g., assembly, insertion, wiping) due to two primary limitations in existing architectures:

Frequency Mismatch: Real-world sensors operate at vastly different sampling rates (e.g., cameras at 15–30 Hz vs. force/torque sensors at 100–1000 Hz). Most existing VLA models fuse all modalities into a single token stream at a unified, low frequency. This necessitates downsampling high-frequency force data, discarding critical reactive cues needed to handle impacts, stick-slip phenomena, and force spikes.
Open-Loop Execution: Standard VLM–Action Expert (AE) pipelines typically execute "action chunks" in an open-loop manner between expensive VLM updates. This lack of high-frequency closed-loop feedback leads to delayed responses to contact changes, causing excessive manipulation forces, task failure, or damage to components.

2. Methodology: FAVLA Framework

The authors propose FAVLA, a Force-Adaptive Fast–Slow VLA model that decouples slow semantic reasoning from fast contact-aware control while maintaining an end-to-end learning scheme.

A. Architecture Design

The model consists of two distinct components operating at different frequencies:

Slow VLM Backbone (Semantic Reasoning):
- Runs at a fixed low frequency.
- Encodes vision (RGB), language instructions, and a history of force measurements.
- Produces a latent contextual representation (KV Cache) and predicts near-future force variations (volatility).
- Uses a Temporal Convolutional Network (TCN) to tokenize historical force data before feeding it into the transformer.
Fast Action Expert (Reactive Control):
- Runs at a variable, high frequency.
- Receives the KV Cache from the Slow VLM and the latest high-frequency force data and proprioceptive state.
- Force Adapter: Instead of concatenating force tokens (which dilutes sparse cues), the model injects force features directly into multiple layers of the Action Expert via cross-attention. This allows the AE to refine action chunks based on real-time contact signals.
- Generates action trajectories using Conditional Flow Matching.

B. Force-Adaptive Inference Strategy

To optimize responsiveness without unnecessary computation, the execution frequency of the Fast Action Expert is dynamically scheduled:

Prediction: The Slow VLM includes a Force Variance Head that predicts the volatility of future contact forces.
Adaptive Scheduling: Based on the predicted variance ( $\tilde{\nu}_t$ $\tilde{ν}_{t}$ ), the system calculates the number of AE executions ( $n_t$ $n_{t}$ ) per VLM step.
- Low Variance (Free-space): The AE runs at a lower frequency ( $n=1$ ) to save compute.
- High Variance (Imminent/Active Contact): The AE runs at a higher frequency ( $n > 1$ ) to enable rapid, closed-loop corrections.
Consistency: To prevent jerky motions caused by stochastic variations in repeated AE inferences, the system fixes the sampled noise across calls within a visual cycle and applies Temporal Ensemble to aggregate overlapping action chunks.

3. Key Contributions

Force-Adaptive Fast–Slow Framework: A novel VLA architecture that explicitly separates slow semantic processing from fast force-reactive control, addressing the sensor frequency mismatch inherent in robotic manipulation.
Force-Injected Action Expert: A design that injects high-frequency force features into multiple layers of the Action Expert via a dedicated adapter, enabling direct, layer-wise force conditioning rather than simple token concatenation.
Adaptive Inference Strategy: A mechanism where the VLM predicts future force volatility to dynamically schedule the Action Expert's execution frequency, balancing reactivity and computational efficiency.
Comprehensive Evaluation: Extensive real-world experiments demonstrating superior performance in high-precision industrial tasks compared to state-of-the-art baselines.

4. Experimental Results

The authors evaluated FAVLA on four real-world contact-rich tasks: USB Insertion, Gear Assembly, Box Flipping, and Board Wiping, using a dual-arm Monte robot with 6-axis force sensors.

Success Rate: FAVLA achieved an average success rate of 80.8%, outperforming the strongest baseline (ForceVLA) by 13.8% and the vision-only baseline ( $\pi_0$ $π_{0}$ ) by 38.0%.
- Notable gains were seen in Gear Assembly (93.3%) and Board Wiping (70.0%), tasks highly sensitive to force regulation.
Force Control: FAVLA significantly reduced peak contact forces, a critical metric for safety and component integrity.
- Gear Assembly: Reduced peak force from 12.0 N (baseline) to 7.7 N.
- Box Flipping: Reduced peak force from 12.2 N (baseline) to 9.9 N.
Ablation Studies:
- Adding the Force-Injected AE improved success rates significantly over vision-only baselines.
- The Force Variance Prediction and Adaptive Frequency mechanisms provided the largest marginal gains, proving that dynamically adjusting inference frequency based on predicted contact dynamics is superior to static high-frequency execution.

5. Significance

FAVLA addresses a critical bottleneck in deploying VLA models for industrial robotics: the inability to react quickly to physical contact. By respecting the natural "fast-slow" nature of robotic sensing (semantic understanding is slow; physical reaction must be fast), FAVLA enables:

Safer Manipulation: Drastic reduction in peak contact forces prevents damage to delicate parts and the robot itself.
Higher Precision: Enables successful completion of tight-tolerance assembly tasks that fail under open-loop or unified-frequency control.
Efficiency: The adaptive strategy avoids wasting computational resources on high-frequency control during non-contact phases, making the system more practical for real-time deployment.

This work establishes a new paradigm for contact-rich manipulation, moving beyond simple multi-modal fusion toward frequency-aware, reactive control architectures.