ForceVLA2: Unleashing Hybrid Force-Position Control with Force Awareness for Contact-Rich Manipulation
Authors: Yang Li, Zhaxizhuoma, Hongru Jiang, Junjie Xia, Hongquan Zhang, Jinda Du, Yunsong Zhou, Jia Zeng, Ce Hao, Jieji Ren, Qiaojun Yu, Cewu Lu, Yu Qiao, Jiangmiao Pang
Organizations: Tongji University · Shanghai AI Laboratory · Lumos Robotics · Shanghai Jiao Tong University · East China Normal University · Shanghai Innovation Institute · Zhongguancun Academy · National University of Singapore · Noematrix Intelligence
Embodied intelligence for contact-rich manipulation has predominantly relied on position control, while explicit awareness and regulation of interaction forces remain under-explored, limiting stability, precision, and robustness in real-world tasks. We propose ForceVLA2, an end-to-end vision-language-action framework that equips robots with hybrid force-position control and explicit force awareness. ForceVLA2 introduces force-based prompts into the VLM expert to construct force-aware task concepts across stages, and employs a Cross-Scale Mixture-of-Experts (MoE) in the action expert to adaptively fuse these concepts with real-time interaction forces for closed-loop hybrid force-position regulation. To support learning and evaluation, we construct ForceVLA2-Dataset, containing 1,000 trajectories over 5 contact-rich tasks, including wiping, pressing, and assembling, with multi-view images, task prompts, proprioceptive state, and force signals. Extensive experiments show that ForceVLA2 substantially improves success rates and reliability in contact-rich manipulation, outperforming pi0 and pi0.5 by 48.0% and 35.0%, respectively, across the 5 tasks, and mitigating common failure modes such as arm overload and unstable contact, thereby actively advancing force-aware interactive physical intelligence in VLAs. The project page is available at https://sites.google.com/view/force-vla2/home.
Figures & tables
Figure 2: Framework of ForceVLA2. ForceVLA2 takes multi-view images, task and force prompts, and proprioceptive states (EE pose and force) as input. Force is injected at multiple scales: as sub-task prompts fused with images in the vision-language model, and as force tokens combined with EE pose in the multimodal encoder, with a bypass to preserve raw signals. The cross-scale MoE integrates these modalities to produce hybrid force–position actions and track sub-task progress for adaptive, contact-rich manipulation.
Figure 3: The illustration of ForceVLA2-Dataset. (a) ForceVLA2-Dataset is the first dataset with force prompts for task decomposition and the only one providing force-control supervision. (b) It features 1,000 demonstrations across five contact-rich tasks.
Figure 4: The dataset collection system. A Flexiv arm is driven by manually controlled GELLO Wu et al. (2023) to accomplish dexterous tasks and record images, force, as well as the pose of the robot.
Figure 5: Qualitative results on typical manipulation tasks (compared with ForceVLA and π serials) . ForceVLA2 completes these tasks with higher success rates and faster execution while avoiding arm overload, demonstrating superior compliance.
Figure 6: Additional tests on following and re-targeting. ForceVLA2 exhibits robust position and orientation following, and in object search tasks, it can still perform successful re-grasps even when visual observations fail.
Type
Method
Press bottle
Clean vase
Clean board
Retri. plate
Assem. gears
Avg.
w/o Force
π0
35.0
20.0
35.0
0.0
0.0
18.0
π0.5
45.0
30.0
45.0
15.0
20.0
31.0
w/ Force
ACP
25.0
30.0
25.0
0.0
0.0
16.0
π0 w/ F
30.0
25.0
20.0
10.0
0.0
17.0
ForceVLA
70.0
25.0
55.0
15.0
10.0
35.0
ForceVLA2
80.0
75.0
70.0
35.0
70.0 ↑50
66.0
Table 1: Success rates ( % ) comparison of different methods. π0 w/ F: π0 with native force input. Each value represents the success rate over 20 trials. Avg. : the mean performance across all five contact-rich tasks. cyan : main metric. bold : best results.
FP
ME
CM
Press bottle
Clean vase
Clean board
Retri. plate
Assem. gears
Avg.
35.0
20.0
35.0
0.0
0.0
18.0
✓
60.0
25.0
40.0
5.0
5.0
27.0 ↑9
✓
✓
60.0
40.0
65.0
5.0
30.0
40.0 ↑13
✓
✓
✓
80.0
75.0
70.0
35.0
70.0
66.0 ↑26
Table 2: Ablation Study of Different Modules in the ForceVLA2 Model. FP : Force Prompt. ME : Multimodal Encoder. CM : Cross-Scale MoE module. Entries indicate success rate ( % ). gray : baseline results.
VM
FM
Press bottle
Clean vase
Clean board
Retri. plate
Assem. gears
Avg.
70.0
30.0
60.0
0.0
20.0
36.0 \penalty
✓
85.0
40.0
85.0
0.0
40.0
50.0 ↑14
✓
✓
80.0
75.0
70.0
35.0
70.0
66.0 ↑16
Table 3: Ablation of input modality combinations in the Cross-Scale MoE module. VM : Visual Modality. FM : Force Modality. Entries indicate success rate ( % ).
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 7: Statistical details of ForceVLA2-Dataset. Frequency distributions of force components (a) and (b) torque components of x , y , z dimensions recorded during task execution. (c) Illustration of short-horizon reactive manipulation skills based on force and position. (d) Skill distribution.
Figure 8: Ablation study of force injection in the multimodal encoder. The figure illustrates three candidate insertion points for the force branch: (1) fusion in the VLM pathway, (2) fusion in the multimodal encoder, and (3) fusion with the state, building upon (2).
Parameter
Value
Learning Rate Schedule
Cosine Decay
Optimizer
AdamW
EMA Decay
0.99
Random Seed
42
Batch Size
32
Training Steps
30,000
Appendix
Table 4: Training details of Foca-VLA.
Method
Press bottle
Clean vase
Clean board
Retri. plate
Assem. gears
Avg.
VLM Pathway
10.0
10.0
5.0
0.0
0.0
5.0
ME
75.0
55.0
65.0
40.0
55.0
58.0
State Fusion
80.0
75.0
70.0
35.0
70.0
66.0
Appendix
Table 5: Success rates ( % ) of kinds of force injection in the multimodal encoder. ME : Multimodal Encoder. State fusion means fusion with state modality, building upon ME.
Contact-rich robotic manipulation studies how robots plan and control physical interactions such as grasping, insertion, pushing, and dexterous object handling.
Humanoid robots are expected to perform diverse human-level tasks in daily environments, many of which require precise regulation of interaction forces. While recent vision-language-action (VLA) models have shown promise for semantic planning and visuomotor control, existing humanoid systems primarily represent actions through geometric motion goals and rely on whole-body controllers focused on motion tracking, with limited explicit reasoning or control of interaction forces. This limitation is particularly relevant in contact-rich tasks, where geometrically similar motions may require different force regimes depending on the task context and where visual observations may become unreliable after contact. In this work, we present Opt2VLA, a force-aware VLA framework that introduces explicit force commands at the VLA-to-control interface for humanoid whole-body manipulation. A single multi-task VLA policy jointly predicts both geometric motion goals and continuous contact-force references, which are tracked by task-specific reinforcement learning (RL)-based whole-body controllers. To provide scalable and physically grounded supervision, we generate dynamically feasible and contact-consistent training data via whole-body trajectory optimization (TO) with explicit force references. We evaluate Opt2VLA on three contact-rich humanoid tasks and show that explicit force conditioning enables more accurate and consistent force regulation than motion-only control, while physically grounded torque supervision from TO further improves force tracking accuracy and stability. Closed-loop evaluations further demonstrate language-conditioned force modulation with Opt2VLA in simulation and on humanoid hardware.
Fukang Liu, Yipu Chen, Jaehwi Jang +3
Institute for Robotics and Intelligent Machines, Georgia Institute of Technology, Atlanta, GA, USA
Vision-Language-Action (VLA) models have demonstrated strong performance across diverse robotic manipulation tasks, yet their predominantly vision-centric perception and position-controlled execution remain insufficient for contact-rich manipulation. Visual observations alone often provide limited evidence of contact onset and interaction magnitude, while position-control policies cannot respond compliantly to rapidly changing contact dynamics. To bridge both the perception and control gaps, we propose TAO-Force, a force-conditioned VLA framework that combines force-aware policy learning with contact-regulated execution. For force-aware perception, TAO-Force introduces Force-conditioned Feature-wise Linear Modulation (F-FiLM) to inject encoded force feedback into the representations of a frozen pretrained visual-language backbone while preserving its semantic priors. For responsive control, it employs a contact-gated fast-slow architecture, with a slow position-control branch tracking nominal trajectories during non-contact phases and a fast admittance-control branch regulating physical interaction during contact phases. Detailed analyses on a force-perception task and real-world evaluations across four contact-rich manipulation tasks validate the effectiveness and robustness of TAO-Force.
Bohan Gan, Xuanzhang Wen, Yongsheng Zhao +8
China Mobile (Hangzhou) Information Technology Co., Ltd., Hangzhou 310023, China
Contact-rich manipulation demands human-like integration of perception and force feedback: vision should guide task progress, while high-frequency interaction control must stabilize contact under uncertainty. Existing learning-based policies often entangle these roles in a monolithic network, trading off global generalization against stable local refinement, while control-centric approaches typically assume a known task structure or learn only controller parameters rather than the structure itself. In this paper, we formalize a physically grounded interaction frame, an instantaneous local basis that decouples force regulation from motion execution, and propose a method to recover it from demonstrations. Based on this, we address both issues by proposing Force Policy, a global-local vision-force policy in which a global policy guides free-space actions using vision, and upon contact, a high-frequency local policy with force feedback estimates the interaction frame and executes hybrid force-position control for stable interaction. Real-world experiments across diverse contact-rich tasks show consistent gains over strong baselines, with more robust contact establishment, more accurate force regulation, and reliable generalization to novel objects with varied geometries and physical properties, ultimately improving both contact stability and execution quality. Project page: https://force-policy.github.io/
Hongjie Fang, Shirun Tang, Mingyu Mei +9
Noematrix · Shanghai Jiao Tong University · Zhejiang University +2