Research map / Vision-language and multimodal / Robot action and navigation
ForceVLA: Enhancing VLA Models with a Force-aware MoE for Contact-rich Manipulation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes ForceVLA, which adds a force-aware Mixture-of-Experts fusion module (FVLMoE) after the VLM in a pi0-based VLA to fuse 6-axis force/torque with vision-language features, and releases a 244-trajectory contact-rich dataset (ForceVLA-Data).
Problem
VLA models rely mainly on vision and language and struggle with contact-rich manipulation requiring force feedback, especially under occlusion or dynamic uncertainty.
Approach
Treat 6-axis external force as a first-class modality: project it into a token appended to VLM outputs, fuse via a top-1 sparse MoE layer after a transformer encoder, and inject the fused tokens additively into the flow-matching action expert.
Key results
- ForceVLA reaches 60.5% average success across five real-world contact-rich tasks, a 23.2-point gain over pi0-base w/ F (37.3%, as stated in the text).
- On the Visual Occlusion setting ForceVLA achieves 90% success versus 60% for pi0-base w/o F.
- In the ablation, ForceVLA reaches 80% success versus 60% for concatenating force after the VLM and 0% for MoE before the VLM.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add force-aware FVLMoE fusion instead of no force input | π0-base · Overall, five contact-rich tasks | average success rate: 37.3% -> 60.5% | Helped |
| add raw force signals directly to state inputs | π0-base · Overall, five contact-rich tasks | average success rate: 37.3% to 40.2% | Helped |
| use FVLMoE fusion rather than direct force input | π0-base · Overall, five contact-rich tasks | average success rate: 40.2% -> 60.5% | Helped |
| add force-aware FVLMoE rather than no force input | π0-base · Cucumber peeling | average peel length: 10.27 -> 14.12 | Helped |
| add force-aware FVLMoE rather than no force input | π0-base · Cucumber peeling | minimum strokes to clean: 14 -> 7 | Helped |
| use ForceVLA rather than the strongest reported baseline | robot policy · Object Gen. 2 | success rate: 35.00% -> 40.00% | Helped |
| use ForceVLA rather than the strongest reported baseline | robot policy · Visual Occlusion | success rate: 60.00% -> 90.00% | Helped |
| use ForceVLA rather than the strongest reported baseline | robot policy · Unstable Socket | success rate: 30.00% -> 20.00% | Hurt |
| add direct force input | π0-fast · Overall, five contact-rich tasks | average success rate: 31.0% to 14.2% | Hurt |
| add direct force input | π0-base · Multi-task joint training | average success rate: 5.0% -> 42.5% | Helped |
| replace π0-base w/ F with ForceVLA | multi-task robot policy · Multi-task joint training | average success rate: 42.5% -> 67.5% | Helped |
| inject force using a linear layer before the VLM rather than the baseline configuration (ablation) | ForceVLA · Force-fusion ablation | success rate: 45% -> 55% | Helped |
Surprising findings
- Adding MoE fusion before the VLM drops success to 0%, indicating that altering a pretrained VLM's input representation disrupts it.
- Naively adding raw force input degrades pi0-fast from 31.0% to 14.2% average success.
Limitations
- Uses estimated external wrench rather than direct high-fidelity force measurements.
- Validated mainly on platforms with integrated, high-cost force-torque sensing.
- Small number of evaluation trials per task and per setting, with some settings using only 10 trials.
Experimental setup
Datasets: ForceVLA-Data
Benchmarks: ForceVLA-Data real-world tasks (bottle pumping, plug insertion, USB insertion, whiteboard wiping, cucumber peeling)
Metrics: task success rate average peel length minimum strokes to clean
Baselines: pi0-base w/o F pi0-base w/ F pi0-fast w/o F pi0-fast w/ F concatenate after VLM linear before VLM MoE before VLM
Builds on: pi0 Mixture-of-Experts PaliGemma SigLIP flow matching
Closest papers by meaning
- ActionUNet: Improving Robustness of VLA Models with Efficient Multi-scale Fine-tuning (NeurIPS 2026 · similarity 0.91)
- ForceFlow: Learning to Feel and Act via Contact-Driven Flow Matching (NeurIPS 2026 · similarity 0.90)
- ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.90)
- VersaCamVLA: Camera-Configurable VLA Policies for Robotic Manipulation (NeurIPS 2026 · similarity 0.90)
- BridgeVLA: Input-Output Alignment for Efficient 3D Manipulation Learning with Vision-Language Models (NeurIPS 2025 · similarity 0.90)
- PriorVLA: Prior-Preserving Adaptation for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.