Manipulation robots face a timing trade-off: they need to explore multiple approach options before contact, then react in milliseconds once they touch an object. FA-RDP, a new visual-force diffusion policy, switches between slow multimodal planning and fast one-step reactive control depending on the task phase, delivering a 30.0 percentage point success boost over strong baselines on real contact-rich tasks.
What Did the Researchers Build?
FA-RDP (Frequency-Adaptive Reactive Diffusion Policy) is an end-to-end policy for contact-rich manipulation tasks such as flipping a box against a fixture, toggling a switch, and pressing a button. Its core idea is that a robot should not run its policy at a single fixed control frequency for an entire episode. Before contact, the robot needs a distribution over possible approach motions to stay collision-free and choose a feasible strategy. After contact, that diversity matters less than reacting quickly to force feedback.
The system has three components: a shared visual-force diffusion backbone that outputs both low- and high-frequency action chunks using frequency-aware positional encoding, a learned multimodality indicator that estimates how ambiguous the current pre-contact behavior is, and manifold consistency distillation that compresses the high-frequency sampler into one-step inference. The same Transformer backbone is reused at both speeds, so no separate models are required. A slow-fast attention structure processes visual context at a lower rate and force feedback at a high rate, with causal masking to prevent future contact information from leaking into predictions.
What Were the Key Results?
Across three real-world tasks — dual box flipping, dual switch toggling, and dual button pressing — FA-RDP achieved the highest average success rate, beating the leading fixed-frequency baseline by 30.0 percentage points. The gains came from a phase-aware strategy: the policy preserved diverse pre-contact action modes that vision-only and regression baselines collapsed, then activated the distilled high-frequency sampler in low-multimodality contact phases.
Component studies confirmed each design choice. Removing multimodality-guided selection and running a fixed high-frequency policy hurt success because the robot lost pre-contact options. Replacing manifold consistency distillation with existing one-step methods — MeanFlow or Consistency Policy — also degraded performance, showing that predicting action chunks on the robot action manifold is important for stable one-step control. Inference latency stayed below 30 ms for high-frequency execution and below 50 ms for low-frequency execution on a consumer CPU-GPU setup.
| Reported Result | Value |
|---|---|
| Average success rate vs. leading baseline | +30.0 percentage points |
| High-frequency inference latency | < 30 ms |
| Low-frequency inference latency | < 50 ms |
| Demonstrations per task | 60 |
| Evaluation trials per task | 20 |
How Does FA-RDP Work?
FA-RDP builds on the consistent reactive diffusion inference mechanism from earlier visual-force diffusion policies. Observations are split into slow tokens, which carry visual-proprioceptive context, and fast tokens, which carry dense force data from the end-effector sensor. A Transformer with a causal action-force mask lets each action token attend only to current and past force tokens; this prevents the network from learning future contact events while still allowing parallel diffusion training.
The frequency adaptation itself is trained in two stages. Stage 1 trains the shared multi-frequency backbone with frequency-aware positional encoding, so one network can predict both a multi-step low-frequency action chunk and a high-frequency chunk. Stage 2 freezes the backbone and trains a small "multimodality indicator" head on a calibration set. For each calibration example, the system samples the low-frequency policy multiple times with different diffusion noise and measures the empirical residual between sampled action chunks. A high residual means many plausible motion options remain; a low residual means the contact phase has narrowed the solution space. During deployment, the indicator reads the current slow tokens and applies a fixed threshold to decide between the low-frequency sampler and the high-frequency distilled sampler.
Because the high-frequency sampler would still require multiple denoising steps, the authors distill it with manifold consistency distillation. Instead of predicting noise-like epsilon, score, or velocity targets, the network is reparameterized to directly predict action chunks on the robot action manifold, while retaining DDPM-based residual supervision. This makes one-step inference stable enough for force-responsive control. A shared 100 Hz command layer also interpolates policy outputs and adjusts each command using the latest measured external force before targets reach the Flexiv Cartesian motion-force interface.
Why Does This Matter for Robotics?
Contact-rich manipulation is where many robot deployments still fail. Assembly, switch operation, plug insertion, and button actuation all require the robot to reason about multiple possible approach paths, then suddenly switch into a stiff, reactive mode once contact is made. FA-RDP's phase-aware control could make these systems more reliable without requiring heavier compute or faster actuators.
The frequency-adaptive design also suggests a practical pattern: spend inference budget where ambiguity is high, and use distilled one-step inference only when the task demands rapid closed-loop response. That pattern applies directly to robot arms on integration lines, humanoid platforms manipulating objects, and warehouse robots performing structured tasks. For teams choosing hardware, browse humanoid robots on Robot Overflow and used industrial robots increasingly need this kind of force-aware, phase-adaptive control to handle real contact tasks.
What Are the Limitations and Open Questions?
FA-RDP was tested on three tasks with only 60 demonstrations each, and the multimodality indicator relies on a fixed threshold tuned on a calibration set. That threshold may need re-tuning when switching tasks, objects, or controller gains. The evaluation also shares a simplified control layer: the 100 Hz force compensation handles translation, but rotational contact errors are not explicitly corrected.
The paper leaves open whether frequency adaptation scales to richer 6-DOF contact tasks, longer horizons, or sim-to-real transfer, and whether the threshold can be learned online rather than fixed. Deploying the policy on edge robotics platforms with tighter compute budgets also still needs validation.
Frequently Asked Questions
Why is a fixed control frequency a problem for contact-rich manipulation?
The robot needs a wide range of possible motions before contact to handle ambiguity, but after contact it needs fast, smooth reactions. A single fixed frequency sacrifices one of those two modes.
How does the multimodality indicator decide which sampler to use?
It is a small head trained on a calibration set. It samples the low-frequency policy multiple times with different noise and measures how much the predicted action chunks vary; high variance means high multimodality and keeps the slow sampler active.
What is manifold consistency distillation?
It is a distillation method that reparameterizes the diffusion network to directly predict action chunks on the robot action manifold instead of noise-like targets. This allows stable one-step high-frequency inference with DDPM residual supervision.
Does FA-RDP need separate models for slow and fast control?
No. A single visual-force diffusion backbone is shared across frequencies using frequency-aware positional encoding, so one network predicts both low-frequency and high-frequency action chunks.
Conclusion
FA-RDP shows that sampling speed and strategy should change with the manipulation phase rather than staying constant. By preserving multimodal pre-contact planning and switching to distilled one-step reactive control under contact, it achieves a significant success improvement on real robots. That combination of diversity and reactivity is exactly what contact-rich tasks demand.