Flow with the Force Field: Learning 3D Compliant Flow Matching Policies from Force and Demonstration-Guided Simulation Data

1. Problem Statement

Robotic manipulation tasks involving continuous contact (e.g., flipping blocks, wiping surfaces, bi-manual assembly) require a delicate balance between precise motion control and compliant force regulation. While modern visuomotor policies (mapping raw visual data to motor actions) have advanced, they often ignore physical compliance, leading to:

• Excessive contact forces and fragile behavior under uncertainty.
• Failure in tasks where fixed-stiffness controllers cannot handle varying contact conditions.
• High data dependency: Existing methods that do incorporate force (e.g., via diffusion models or reinforcement learning) typically require hundreds of real-world demonstrations or complex reward engineering, limiting scalability.
• Sim2Real Gap: Generating high-quality simulation data for contact-rich tasks is computationally expensive, and standard simulation data often fails to transfer to real robots without significant domain adaptation.

The core challenge addressed is: How to learn adaptive, compliant visuomotor policies from a single human demonstration in simulation, enabling zero-shot transfer to real robots for contact-rich tasks without real-world data collection.

2. Methodology

The authors propose a three-stage framework: Force-Informed Data Generation, Compliant Flow Matching Policy Learning, and Safe Real-World Rollout.

A. Force-Informed Data Generation (Simulation)

Instead of collecting massive real-world datasets, the system generates diverse training data in simulation (IsaacGym) from a single teleoperated demonstration ($D_h$).

• Trajectory Decomposition: The demonstration is split into free-space ($D_f$) and in-contact ($D_c$) segments.
• Domain Randomization: Initial conditions (object pose, mass, friction, end-effector pose) are randomized to create new scenarios.
• Trajectory Warping:
  • Free-Space: Uses Laplacian Editing to warp the trajectory to new initial conditions while preserving local geometric properties.
  • In-Contact: Uses an object-centric frame to warp poses and forces relative to the randomized object.
  • Force-Informed Virtual Targets: A critical innovation where a "virtual target" is constructed for each contact point by shifting the end-effector pose in the direction of the measured force ($x_{ee}^v = x_{ee} + k_f F$). This forces the policy to learn the intention of maintaining contact rather than just following a path.
• Output: A diverse synthetic dataset ($D_{sim}$) containing point clouds, end-effector poses, and force measurements.

B. Adaptive Compliant Flow Matching Policy

The policy is trained using Flow Matching, a generative modeling technique that offers higher inference frequencies than diffusion models, making it suitable for high-frequency control.

• Inputs:
  • Point Clouds ($o_{pc}$): Single-view 3D point clouds (processed via DP3 Encoder) for spatial generalization, replacing RGB images.
  • End-Effector Pose ($o_{ee}$) and Force ($o_f$): Processed via MLPs.
• Outputs: The policy predicts a trajectory horizon ($H=16$) consisting of:
  1. Reference Pose ($X_{ref}$): The desired trajectory.
  2. Virtual Target ($X_{vt}$): The force-informed target used for compliance.
  3. Impedance Parameter ($d$): A scalar gain controlling compliance, scheduled based on estimated force magnitude.
• Training Objective: The model learns to match the velocity field of the interpolation path between noise and the target action distribution, conditioned on the sensor inputs.

C. Compliant Rollout with State-Vector Fields

To ensure safety and robustness during real-world execution, the policy output is not executed via rigid position tracking.

• State-Velocity Field: The policy's pose output is converted into a desired velocity field.
• Passive Impedance Controller: The robot is controlled by a passive impedance law: $F_c = D(x)(\dot{x} - f(x))$.
  • Instead of rigidly tracking waypoints, the controller blends the reference motion direction with the compliant direction (vector from current position to the virtual target).
  • The blending gain ($d$) is dynamically adjusted based on contact force: high force $\rightarrow$ high compliance (lower gain); low force $\rightarrow$ low compliance (higher tracking fidelity).
  • This approach injects less energy into the environment and allows the robot to "yield" naturally if it encounters unexpected obstacles, preventing damage.

3. Key Contributions

1. Lightweight Data Generation Strategy: A novel pipeline to generate force-informed, diverse simulation data from a single human demonstration using Laplacian editing and force-modulated virtual targets, eliminating the need for extensive real-world data collection.
2. Adaptive Compliant Flow Matching Policy: A new policy architecture that takes 3D point clouds and force as inputs and outputs both pose actions and impedance parameters, enabling high-frequency, contact-aware control.
3. Zero-Shot Sim2Real Transfer: Successful deployment on real Franka robots for Block Flipping and Bi-manual Moving tasks without any real-world demonstrations or fine-tuning.
4. Passive Impedance Rollout Scheme: A control strategy that converts policy poses into a state-velocity field, coupled with a passive impedance controller. This reduces energy injection and improves safety compared to classical position-based impedance controllers.

4. Experimental Results

The framework was validated on two real-world tasks using Franka Emika robots:

• Tasks:
  1. Block Flipping: Pivoting a block from flat to upright (approx. 85°).
  2. Bi-manual Moving: Two arms cooperatively placing a large object onto a platform.
• Generalization Performance:
  • Spatial Generalization: The full method achieved 97.6% success in Block Flipping and 82.9% in Bi-manual Moving across randomized positions, significantly outperforming baselines (which often failed at 0% or <30%).
  • Object Generalization: The policy generalized to unseen object sizes and shapes (e.g., different box dimensions) with high success rates, whereas point-cloud-only or force-only baselines failed to maintain contact or got stuck.
• Ablation Studies:
  • Removing Laplacian editing or force-informed targets dropped success rates to 0%, proving the necessity of diverse data and force-awareness.
  • Removing force input or compliant output significantly degraded performance, confirming that both geometric (point cloud) and physical (force) feedback are essential.
• Controller Comparison:
  • The proposed Passive Impedance Controller achieved a 10/12 success rate with lower mean energy injection (0.734 J) compared to the Classical Impedance Controller (8/12 success, 0.835 J). The classical controller often failed by pushing too hard against the object, causing it to slip or jam.

5. Significance

This work bridges a critical gap in robotic learning by demonstrating that compliance can be learned effectively from simulation without real-world data collection.

• Scalability: It removes the bottleneck of collecting hundreds of real-world demonstrations for contact-rich tasks.
• Safety & Robustness: By integrating force feedback and passive impedance control, the system handles uncertainty and physical constraints more safely than rigid position controllers.
• Generalization: The use of 3D point clouds combined with force data allows the policy to generalize across different object geometries and spatial configurations, a capability often missing in RGB-based or purely kinematic approaches.
• Practicality: The framework provides a complete pipeline from a single simulation demo to a robust real-world deployment, offering a viable path for deploying complex manipulation skills in unstructured environments.