FreqFM, a frequency-conditioned Flow Matching method for vision-language-action (VLA) robots, makes action generation explicitly aware of fast and slow motion patterns. It reshapes the model’s source distribution, training objective, and inference guidance using frequency statistics, improving a paired temporal baseline across LIBERO, LIBERO-Plus, VLA-Arena, and six real-robot tasks without changing the VLA backbone.
What did the researchers build?
The researchers built FreqFM, a Flow Matching action-generation framework for VLA models. VLA systems combine visual observations, language instructions, and robot actions. Instead of predicting each action only in its original time order, FreqFM also represents the action sequence through frequency components.
The method applies a discrete cosine transform (DCT) across the action horizon. A DCT converts a finite sequence into coefficients that describe broad motion trends and faster changes. In practical terms, low-frequency components capture gradual movements such as reaching or sweeping, while higher-frequency components describe quicker adjustments, corrections, or vibrations.
FreqFM estimates a power spectrum from the training actions. This spectrum records how much motion energy appears at each frequency for each robot control dimension, such as a joint, gripper channel, or Cartesian motion coordinate. The resulting statistics become an explicit conditioning signal throughout the Flow Matching pipeline.
The framework changes three stages: how the initial source sample is constructed, how the training objective weights motion components, and how conditional guidance is allocated during inference. The approach is designed as an add-on to an existing VLA backbone rather than a replacement for the visual or language model.
What results did FreqFM achieve?
FreqFM was evaluated against a paired temporal Flow Matching baseline on the LIBERO and LIBERO-Plus simulation benchmarks, the VLA-Arena benchmark, and six real-robot tasks. The reported conclusion is consistent across these settings: making frequency explicit improves performance over the temporal baseline without requiring changes to the underlying VLA architecture.
The comparison matters because the baseline already uses Flow Matching, so the reported improvement comes from how action trajectories are represented and controlled rather than from switching to an unrelated policy family. FreqFM also preserves the same general transport-based generation process, making the result relevant to teams already experimenting with Flow Matching action experts.
The available paper text states that the method improves the baseline by up to a formula-defined number of points across the listed benchmarks and real-robot tasks. The exact maximum gain is not exposed in the supplied extraction, so a precise score difference cannot be reported responsibly. The text does identify six real-robot tasks as part of the evaluation, providing evidence beyond simulation.
VLA-Arena also reveals an important trade-off. FreqFM performs worse on its Extrapolation task, where action statistics are least aligned with the training spectrum. That result shows both the value and the risk of using dataset-level frequency assumptions.
How does frequency-conditioned Flow Matching work?
Flow Matching learns a continuous transport process. It starts with a source sample, typically representing a simple distribution, and learns a velocity field that moves that sample toward a demonstrated action chunk. During inference, an ordinary differential equation follows the learned velocity field until it produces the robot’s action sequence.
FreqFM modifies this process in three connected steps:
- Frequency-aware source construction: Training actions are transformed with the DCT, and a per-frequency, per-control-dimension power spectrum is estimated. The initial source distribution is then given frequency-dependent variance instead of treating every temporal pattern as equally likely. This makes the source resemble the target action statistics before transport begins and reduces the mismatch between source and target covariance.
- Frequency-aware training objective: The transport objective is expressed in frequency coordinates, allowing the model to learn motion components with their own scales. The framework derives a reference transport scale for each frequency. Along the linear transport path, the target velocity at a given frequency has a second moment equal to twice that frequency’s action power, linking the training signal directly to observed motion energy.
- Frequency-aware inference guidance: Conditional guidance helps a generated trajectory follow the visual scene and language instruction. FreqFM budgets the guidance residual by frequency rather than applying one undifferentiated allowance to the entire action sequence. This gives the inference process a principled scale for balancing condition-following against the action statistics learned from the dataset.
The DCT is important because it provides a fixed, interpretable basis for finite action chunks. A frequency-selective operation that would require dense interactions between many timesteps in the original representation becomes a compact set of per-frequency operations. The same spectrum can therefore guide source sampling, optimization, and inference.
Why does FreqFM matter for robotics?
Robot actions are not equally difficult across all timescales. A robot arm may need a slow, coordinated reach, followed by fast wrist corrections and a precise gripper movement. A model that treats the full trajectory as one undifferentiated time sequence can learn these relationships, but it has no explicit control over which motion scales receive attention, variance, or guidance.
FreqFM offers an engineering handle for that problem. Frequency statistics can help preserve smooth large-scale motion while managing faster corrections during generation. That distinction is relevant to manipulation, mobile manipulation, and warehouse tasks where trajectory smoothness affects grasp success, payload stability, wear, and safety.
The method also reduces the integration burden because the VLA backbone remains unchanged. Teams evaluating warehouse robots or used cobots for sale could view frequency-conditioned action generation as a policy-layer improvement rather than a hardware replacement.
The strongest practical signal is the evaluation beyond simulation: six real-robot tasks improved over the paired temporal baseline. However, the Extrapolation result shows that frequency-aware control works best when deployment actions resemble the training distribution.
What limitations remain?
FreqFM relies on fixed, training-set-level spectral statistics. The source distribution and guidance budget do not adapt directly to the current camera observation, language instruction, or task context. That design simplifies the method but can hurt performance when a task requires motion patterns that differ substantially from the training data.
The VLA-Arena Extrapolation regression is the clearest example. It affects tasks whose action statistics are least likely to match the learned spectrum, showing that a dataset average cannot represent every deployment condition.
The method also requires smooth action dimensions to be designated for each robot embodiment. That introduces setup work and makes cross-robot transfer less automatic. The supplied results establish benefits on the reported tasks, but they do not establish how FreqFM behaves across many robot morphologies, highly discontinuous actions, or rapidly changing control modes.
What are the most common questions about FreqFM?
What problem does FreqFM solve?
It gives VLA action generation an explicit representation of fast and slow motion patterns, rather than leaving frequency structure implicit in time-domain processing.
Does FreqFM require a new VLA backbone?
No. The framework conditions the Flow Matching action expert and leaves the underlying VLA backbone unchanged.
What is the DCT doing in the method?
The discrete cosine transform converts an action sequence into frequency components that describe different motion scales, making those components available for modeling and inference control.
Where does FreqFM struggle?
It struggles when deployment actions do not match the training-set spectrum, with the reported Extrapolation regression highlighting that limitation.
What is the conclusion?
FreqFM turns action frequency into an explicit modeling dimension for VLA-based Flow Matching. Its spectrum-aware source construction, training objective, and inference guidance improve a temporal baseline across several benchmarks and six real-robot tasks, while the fixed-spectrum design remains vulnerable to distribution shift.