ForceVLA: from sensing force to using it
What changes when force is fused into action generation, rather than simply added as another input?
INTERACT ROBOTICS · Research Notes · Reviewed September 7, 2026
The research question
A VLA may understand a task through vision and language, while subtle resistance after contact still needs direct observation. ForceVLA proposes a force-aware mixture-of-experts module that connects force and torque to action generation.
Comparisons within the paper
| Condition | Mean success across 5 tasks |
|---|---|
| π0-base / No force input | 37.3% |
| π0-base / Force added directly | 40.2% |
| ForceVLA / FVLMoE fusion | 60.5% |
ForceVLA v3, §5.2. The change from 37.3% to 60.5% is 23.2 percentage points. These are external research results, not our performance or a comparison with π0.7.
A signal needs the right representation.
The proposed fusion method improved performance more than simply adding force input. Having a signal is not enough if the policy does not learn when to use it.
We read this as a reason to design acquisition and learning together. Preserve force direction, frames, and time history, then study how representations should deliver that information to the policy at each contact stage.
Reading the results in context
These results apply to the paper’s robot, data, and five tasks. Changes to baseline force inputs, pretraining, or fine-tuning may change the difference. The same gain cannot be assigned to every VLA, nor can maximum gains from different papers be combined.
References
- ForceVLA: Enhancing VLA Models with a Force-aware MoE for Contact-rich Manipulation ↗
2025 / v3 / §5.2, Figure 5. Experimental conditions must remain attached to the comparison.