ForceVLA2: Unleashing Hybrid Force-Position Control with Force Awareness for Contact-Rich Manipulation

• Current VLAs perform semantic grounding and language following but lack physical-dynamics reasoning for fine-grained contact interactions.Source: §1 · p. 2
• Existing VLAs typically treat force as an auxiliary perceptual input rather than an active signal for adaptive force–position interaction.Source: §1 · p. 2
• ForceVLA2 combines force prompts in the VLM with a Cross-Scale MoE that fuses task concepts and interaction forces for hybrid control.Source: §1 · p. 2
• ForceVLA2-Dataset contains 1,000 trajectories across 5 contact-rich tasks with multimodal observations, force prompts, and force-aware action supervision.Source: §1 · p. 3
• 48.0% and 35.0% success-rate increases over π0 and π0.5, respectively, are reported across the 5 tasks.Source: §1 · p. 3

Vision-language-action models

• VLA approaches map visual observations and language instructions to actions using autoregressive, diffusion-based, or flow-based policy formulations.Source: §2 · p. 3

Physical modality integration in robot policies

• Force-aware work improves contact-rich manipulation, but force is commonly used as a perceptual cue rather than an active control signal.Source: §2 · p. 3
• Tactile methods improve robustness under occlusion but do not replace direct force sensing because tactile-to-force estimation remains indirect and noisy.Source: §2 · p. 3

Multimodal embodied datasets

• Large embodied datasets rarely provide standardized force modalities, producing sparse and inconsistent coverage for contact modeling.Source: §2 · p. 3
• RH20T and REASSEMBLE add multimodal force data for general manipulation and assembly, respectively, including synchronized visual, proprioceptive, and force signals.Source: §2 · pp. 3–4

Inputs and multimodal integration

• ForceVLA2 jointly encodes multi-view images, task prompts, force prompts, EE pose, and force observations for contact-rich manipulation.Source: Figure 2 · p. 4, §3 · p. 4
• Force prompts provide subtask context for high-level task sequencing and spatial–semantic reasoning when visual cues are insufficient.Source: §3.1 · p. 5

Proprioceptive-force encoding

• A direct force pathway bypasses high-level fusion, preserving transient signals for short-term reactive manipulation alongside multimodal state fusion.Source: §3.2 · p. 5, §3.2.1 · p. 6

Adaptive routing and decoding

• The Cross-Scale MoE uses visual, state, and force experts with dynamic token-wise routing to emphasize the most informative modality at each phase.Source: §3.2.2 · p. 6
• Flow matching progressively denoises actions conditioned on fused visual–language–force context to produce desired pose changes and contact forces.Source: §3.2.2 · pp. 6–7
• The transition indicator combines orientation, distance, and force conditions; exceeding its threshold advances the controller to the next subtask.Source: §3.2.2 · p. 7

Collection and task design

• The dataset uses force-feedback GELLO teleoperation to collect demonstrations with natural motion dynamics and explicitly recorded force-control signals.Source: §4 · p. 8
• Demonstrations combine three RGB camera views, synchronized EE poses and joint states, and 6D force/torque measurements recorded at 300 Hz.Source: §4 · p. 8
• The dataset covers five tasks: press bottle, clean vase, clean board, retrieve plate, and assemble gears, with force-defined subtask boundaries.Source: §4 · p. 8

Dataset statistics

• 1,000 trajectories contain approximately 500K synchronized timesteps of RGB images, EE poses, and force signals.Source: §4 · p. 8
• Force prompts and force-integrated actions support subtask-structured policy learning for embodied hybrid force–position interaction.Source: §4 · p. 8

Experiment Setting

• The benchmark evaluates five contact-rich tasks using success rate over 20 independent trials per task.Tasks include pressing, wiping, retrieving, and assembling.Source: §5.1 · p. 9

Main Results

• ForceVLA2 achieves a 66% average success rate across all tasks and surpasses the second-best model by up to 50% on force-sensitive tasks.The comparison includes ACP, π0, π0.5, ForceVLA, and π0 with force input.Source: §5.2 · p. 10, §5.1 · p. 10
• ForceVLA2 actively adjusts interaction forces in force-intensive tasks, avoiding overload where other VLAs often fail.The reported tasks include clean board and bottle pressing.Source: §5.2 · p. 10
• ForceVLA2 rapidly tracks force changes after sudden perturbations, while other VLAs slowly follow the new end-effector pose and lose stable contact.In bottle pressing, lowering the base changes the contact configuration during execution.Source: §5.2 · p. 10
• ForceVLA2 maintains contact during misaligned vase wiping and retries failed grasps during object search through force-guided probing.These behaviors contrast with trajectory continuation, overload, and failure reported for other VLAs.Source: §5.2 · p. 11

Ablation Study

• Success rates increase consistently as Force Prompt, Cross-Scale MoE, and Multimodal Encoder modules are added, with Cross-Scale MoE providing the largest gain.The ablation progressively adds modules on top of π0 across five tasks.Source: §5.3 · p. 11
• Visual and force inputs contribute roughly equally in Cross-Scale MoE, while an extra force token can slightly reduce performance on Clean Board.The reduction is attributed to simple force requirements where additional degrees of freedom may perturb a stable policy.Source: §5.3 · pp. 11–12

Conclusion

• ForceVLA2 achieves 66% average success across five tasks, compared with 48% for π0, 35% for π0.5, and 31% for ForceVLA.It also surpasses the second-best method by 50 percentage points on assembling gears.Source: §6 · p. 12
• ForceVLA2 reduces failures caused by arm overload and unstable contact while force prompts, Cross-Scale MoE, and multimodal fusion improve performance monotonically.The full ForceVLA2 variant achieves the best ablation results.Source: §6 · p. 12

• A direct reactive force pathway improves observability and controllability, enabling rapid feedback, reduced reliance on past trajectories, and active force-guided interaction.Source: A. Theoretical Analysis of Force Pathway · pp. 13–14
• Position-only control restricts the reachable task space because force is determined by unknown environment dynamics rather than independently controlled.Source: A. Theoretical Analysis of Force Pathway · p. 13
• The position-only reachable set forms a 6-dimensional manifold embedded in the 12-dimensional task space.Source: A. Theoretical Analysis of Force Pathway · p. 13
• ForceVLA2 uses a 12-dimensional hybrid force–position action policy that learns from interaction observations and coordinates both command types.Source: A. Theoretical Analysis of Force Pathway · p. 14
• The effective controllability index is 6/12 = 0.5 for position-only control, while the force-aware policy approaches a task-dependent maximum as training data and physical constraints permit.Source: A. Theoretical Analysis of Force Pathway · p. 14

• Force distributions differ by task: pressing concentrates force along negative z, wiping spreads forces across axes, probing centers near zero, and assembly shows broader variation.Source: B. Statistical details of the… · p. 14
• Torque distributions are near zero for pressing but wider for wiping and probing, while gear assembly has the most diverse z-axis torque distribution.Source: B. Statistical details of the… · pp. 14–15
• Figure 7 summarizes force and torque component frequencies, reactive force-position skills, and skill distributions in the dataset.Source: Figure 7 · p. 15
• Five short-horizon skills are identified from force and position trajectories using task-specific displacement and force thresholds.Source: B. Statistical details of the… · p. 15

• Foca-VLA training uses 8 A100 GPUs, batch size 32, 30,000 steps, approximately 10 hours, and reaches 15 Hz inference on a 4090 GPU.Source: C. Implementation and Experiments · p. 15
• The ablation compares force injection in the VLM pathway, the multimodal encoder, and fusion with the EE 6D pose state.Source: C. Implementation and Experiments · p. 16, Figure 8 · p. 16
• Force injection into the VLM pathway substantially degrades performance, while multimodal-encoder and state-fusion integration consistently improve it.Source: C. Implementation and Experiments · p. 16
• The final design incorporates force at both the multimodal encoder and state-fusion levels.Source: C. Implementation and Experiments · p. 16

• Jacobian-based mapping decouples the policy from robot kinematics and supports torque-controlled robots, EE force/torque-sensor robots, and torque-interface actuators.Source: D. Discussion on hardware and evaluation · p. 16
• Real-world evaluation is used because friction and contact modeling make simulation less reliable for force-aware manipulation.Source: D. Discussion on hardware and evaluation · p. 17
• The evaluation also addresses the absence of a widely available benchmark directly usable for force-aware VLAs.Source: D. Discussion on hardware and evaluation · p. 17