CR-VLA-Force: Learning Control-aware Compliance VLA Model for Robust Contact-rich Robotic Manipulation

• Contact-rich tasks remain challenging because small perception or control errors can trigger force transients, collisions, or contact loss and unsafe interactions.Source: §1 · p. 2
• Existing force-aware VLA models remain limited in force tracking and rapid adjustments, motivating compliance priors that combine learning adaptability with safety-aware control.Source: §1 · p. 2
• A multi-stage strategy trains the VLA backbone for perception and planning before using a force-aware MoE to correct actions, with force and pose as prediction targets.Source: §1 · pp. 2–3
• A force-history encoder represents temporal contact cues in addition to real-time force signals for cross-modal fusion.Source: §1 · p. 3
• The VLA-guided adaptive compliance controller uses predicted desired positions and feedforward forces for contact-rich, high-frequency control.Source: §1 · p. 3
• CC-VLA increases success rates and force-position tracking precision while maintaining high performance in partial out-of-distribution scenarios.Source: §1 · p. 3

2.1 End-to-End force/torque aware policy learning

• Diffusion policies and related methods condition on images, pose, and force/torque signals, with some approaches closing the loop through compliance control or handling slow and fast feedback hierarchically.Source: §2.1 · p. 3
• Force-aware VLA models use MoE fusion or force prediction errors to condition pose action chunks, yet remain limited in accurate force tracking and reactive force-position adjustment.Source: §2.1 · p. 3

2.2 Compliance control with impedance learning

• Impedance learning includes demonstrations, online stiffness optimization with passivity constraints, inverse reinforcement learning, and safe reinforcement learning for contact-rich tasks.Source: §2.2 · p. 4
• Despite these advances, closely coupling variable-compliance controllers with control-aware multimodal policies remains an outstanding research challenge.Source: §2.2 · p. 4

3 METHODOLOGY

• The system builds on π0, encoding visual observations and instructions before a flow-based action expert predicts action chunks, with force-aware fusion refining the policy.The model is described as a slow–fast hierarchy combining a multimodal policy and an adaptive compliance controller.Source: §3 · p. 4

3.1 Problem formulation

• The policy conditions action sequences on language, multi-view images, proprioceptive state, real-time force, and a historical force window.Its control objective is for the robot’s actual trajectory to follow an interaction trajectory informed by real-time force.Source: §3.1 · pp. 4–5

3.2 Shared Teleoperation Data Acquisition

• Compensated virtual impedance softens the teleoperation interaction, allowing greater contact displacement for the same force and recording the robot’s proprioceptive state.The method adjusts the reference pose using force feedback during demonstrations.Source: §3.2 · pp. 5–6

3.3 Historical Force Sequence Encoder

• A learned force token summarizes segmented, position-encoded F/T history through cross-attention, producing a compact feature for downstream policy decisions.The encoding aggregates long-horizon interaction cues while avoiding the computational overhead of processing the full force history.Source: §3.3 · p. 6

3.4 MoE-Based Feature Fusion with Multi-Stage Learning Strategy

• Training first aligns visual, language, and proprioceptive action spaces, then adds a sparse-MoE fusion expert as a residual branch to refine trajectories with contact dynamics.The second stage jointly optimizes the shared expert so force-aware adjustments refine the established visual policy.Source: §3.4 · pp. 6–7

3.5 Adaptive Compliance Controller

• An asynchronous trajectory layer interpolates action chunks into higher-frequency commands; the adaptive controller executes desired actions using compliant force tracking.Orientations are interpolated on SO(3) along geodesic paths; the controller uses a heuristic stiffness adaptation rather than QP optimization because predicted actions and forces may fluctuate stochastically.Source: §3.5 · pp. 7–8

4 Experiments

• The experiments compared CC-VLA with Diffusion Policy, Pi0, Pi0 with force concatenated to state inputs, Pi05, and ForceVLA, and ablated the VLA model and adaptive compliance controller.The force-tracking comparison reports errors in Table 1, and the ablation comparison reports them in Table 2; the supplied passages give no table values.Source: §4 · p. 8, Table 1 · p. 11, Table 2 · p. 11

4.1 Experimental Settings

• The system used a position-controlled UR5e, four-expert sparse MoE inference once per second with a 40-step action horizon, force sampling every 0.1 s, and 500 Hz control.Stiffness bounds ranged from 400 N/m to 2500 N/m.Source: §4.1 · p. 8
• Evaluations covered four contact-rich tasks; button pressing, window opening, and board wiping used 50 demonstrations each, while plug insertion used 100, and each test condition comprised 24–25 trials.Button and board placements were 6 cm lower than during data collection; proximal plug starts placed the plug near the socket, where the wrist-mounted camera could not capture the socket holes.Source: §4.1 · pp. 9–10

4.2 Results

• 89.2% average success across six conditions, exceeding DP (31.3%), Pi0 (47.3%), Pi05 (44.7%), Pi0 wi F (60.2%), and ForceVLA (73.2%).Figure 6 reports success rates across methods for three contact-rich tasks.Source: Figure 6 · p. 9, §4.2 · p. 10
• CC-VLA was the only method reported to robustly track the 40 N desired force in board-wiping base and OOD experiments.Pi0 and Pi05 triggered protective stops above 75 N, while DP had insufficient force or intermittent contact loss; Figure 7 omits Pi0 and Pi05.Source: Figure 7 · p. 9, §4.2 · p. 10

4.3 Discussion

• Button-pressing, plug-insertion, and window-opening results support CC-VLA's force-sensing performance and robustness under the tested OOD conditions.Window-opening results also highlight performance in compliant, precise, long-range manipulation.Source: §4.3 · pp. 10–11
• Ablations report a 12% button-press success-rate gain and a false-press pose-rate reduction from 28% to 8% with multi-stage training; plug-insertion success rose by 8% to 76%.For proximal plug starts, the historical force-sequence encoder ablation reports success increasing from 72% to 92%. Board-wiping ablations support the practical value of the compliance controller and variable stiffness.Source: §4.3 · p. 11
• The safety study reports that the variable impedance law maintains compliance under excessive forces and external disturbances, while limiting excessive contact forces from potentially hazardous inferred poses.The study also reports balancing safe interaction forces with task success by learning human action adjustments from demonstrations under excessive contact forces.Source: §4.3 · p. 11

5 Conclusion

• CC-VLA significantly outperforms existing methods on challenging contact-rich tasks and force-tracking performance.The authors also report that it can be applied to fine-tune other pretrained VLA models, indicating scalability.Source: §5 · p. 11
• A control-aware VLA framework with a compliance controller targets force-action adjustment latency and poor force-tracking accuracy through reactive force–position control.The approach combines a two-stage training strategy, a force-history sequence encoder, and a multimodal MoE.Source: §5 · p. 11
• The method relies on precise force control and temporal consistency during teleoperation demonstration acquisition, and focuses exclusively on object-level variable-impedance manipulation.Future work will investigate integrating finger-level gripper control with object-level end-effector motion.Source: §5 · p. 12