TL;DR
ReGround is an inference-time method that allows "frozen" Vision-Language-Action (VLA) models to adapt to new robot hardware without retraining or fine-tuning. By learning a robot's specific dynamics through task-agnostic "motor babbling," ReGround translates the VLA’s intended motion into actions the new body can actually execute, restoring near-perfect performance even under severe latency or mechanical shifts.
What Problem Does This Paper Address?
Vision-Language-Action (VLA) models are the "brains" of modern robotics, mapping pixels and text instructions directly to motor commands. However, these brains are typically trained on specific robot bodies. When you deploy a frozen VLA on a robot with different physical characteristics—such as mechanical wear, different controller gains, or significant command latency—the model often fails.
The authors call this actuation shift. Even if the VLA correctly identifies the task (e.g., "pick up the cup"), the specific motor commands it issues might cause the robot to overshoot, move too slowly, or collide with objects because the "mapping" between command and physical result has changed. Standard solutions involve fine-tuning the model on the new hardware, but this is computationally expensive, requires new task demonstrations, and is impossible if the model weights are proprietary or "frozen."
Methods
The authors propose ReGround, a four-stage pipeline that operates entirely at inference time:
- Intent Read-out: Since the VLA was trained on a "nominal" robot, ReGround uses a simple forward model of that original robot to "read" the VLA’s output and determine the trajectory it intended to follow.
- Deployed-body Identification: To understand the new robot, ReGround performs "motor babbling"—roughly 50 seconds of random, task-agnostic movements. This data is used to fit a forward model ($f_{dep}$) of the new hardware, identifying parameters like latency and action mapping.
- Trust-region Tracking: ReGround uses a receding-horizon controller (Model Predictive Control) to find actions that make the new robot follow the intended trajectory. It includes a "trust region" to ensure the new actions don't stray too far from the VLA's original plan, preserving the model's high-level reasoning.
- Shift Monitor: To ensure ReGround doesn't interfere when the robot is working perfectly, a CUSUM (Cumulative Sum) monitor tracks prediction residuals. It only activates ReGround’s corrections if it detects a statistically significant shift in dynamics.
Datasets
The researchers evaluated ReGround using:
- LIBERO: A benchmark suite for robotic manipulation. They specifically used eight tasks within the MuJoCo/robosuite environment.
- VLA Models: They tested two different architectures: $\pi_{0.5}$ (a vision-language-action flow model) and SmolVLA (an efficient, smaller VLA).
- Shift Settings: They simulated five types of actuation shift: command latency (up to 400ms), action-frame rotation (90°), controller gain changes ($\times3$), and a composite shift combining multiple factors.
Key Insights and Ideas
- Intent is Invariant: The core "aha" moment is that while a VLA’s actions might be wrong for a new body, its intended motion is usually still correct for the task. By treating the VLA as a trajectory generator rather than a direct controller, we can re-map those trajectories to any hardware.
- Task-Agnostic Adaptation: Remarkably, ReGround doesn't need to see the robot performing a specific task to fix it. Random "babbling" is sufficient to learn the underlying physics (the "body"), which then generalizes across all tasks the "brain" knows how to do.
- Safety via Monitoring: Actuation shift isn't just a performance bug; it's a safety hazard. The paper shows that naive VLAs under shift often cause unintended collisions. ReGround significantly reduces these unsafe episodes compared to simple linear calibration.
Key Results
ReGround demonstrated a dramatic ability to recover lost performance:
- Latency Recovery: Under a severe 400ms command latency, the base $\pi_{0.5}$ model's success rate plummeted to 16%. ReGround restored it to 90%.
- Rotation and Gain: In scenarios with 90° frame rotation or triple controller gain—where the base model failed completely (0% success)—ReGround restored success to 96-98%, matching "Oracle" performance (where the shift is known perfectly).
- Efficiency: The system requires only about 50 seconds of robot time (1k transitions) to learn the new dynamics well enough to achieve 98% success.
- Minimal Overhead: The CUSUM monitor had a false-alarm rate of only 2.1% on unshifted bodies, meaning the system stays out of the way until it is actually needed.
Limitations and Future Work
- Intent Validity: ReGround assumes the VLA's intended trajectory is still physically possible and logically sound for the task. If the actuation shift is so extreme that the original strategy no longer works (e.g., a motor is completely dead), ReGround cannot "re-plan" a new strategy.
- State Estimation: The method relies on accurate robot state feedback (joint positions, velocities). In highly noisy real-world environments, the accuracy of the learned dynamics model might degrade.
- Complex Dynamics: While effective for latency and gains, the paper primarily looks at rigid-body shifts. Future work could explore more complex shifts, such as changes in link mass or contact dynamics (e.g., a robot picking up a very heavy tool).
Summary updated