Force feedback provides vision-language-action (VLA) policies with physical observations that can guide action selection and execution. However, differences in tasks, demonstrations, and control settings across existing systems hinder reliable comparisons of force integration designs. We introduce ForceBench, a simulation benchmark for evaluating six-axis wrist-force integration under a unified protocol. ForceBench combines eight tasks spanning physical-property inference, resistance-dependent interaction, and contact-rich assembly with 1,600 synchronized demonstrations and task-specific success criteria. We compare one no-force baseline and six force integration designs on a shared π0.5 backbone. Prompt Fusion achieves the highest observed mean simulation success of 35.31%, compared with 23.25% for the baseline, and improves success rates across all three task categories. Predictive and temporal gains in simulation concentrate on physical-property inference, while both temporal designs underperform the baseline on assembly. Real-world evaluation of eight systems across five tasks supports the main qualitative trends: Prompt Fusion and Dual-Path Fusion achieve 63.00% mean success, compared with 40.00% for the baseline, but only Prompt Fusion improves every task. These findings support ForceBench's practical value for identifying the benefits and limitations of force integration across manipulation demands. The project page is available at https://forcebench-platform.github.io/.
Figures & tables
Fig. 2: Simulation tasks. (a) Weight Sort and (b) CoM Sort require physical-property inference; (c) Nut Thread requires contact-rich assembly; (d) Faucet Turn and Faucet Direction, (e) Stove Turn and Stove Press, and (f) Drawer Pull require resistance-dependent interaction. Insets mark the hidden property or resistance transition that appearance does not reveal; Faucet Direction and Stove Press share the faucet and stove environments, respectively.
Fig. 3: Force integration on a shared π0.5 backbone. 1 – 4 denote Cross-Attention, Token Fusion + Pred, Prompt Fusion, and Token Fusion, respectively; π0.5 (No Force) denotes the baseline. 5 and 6 show the additional Temporal Prompt and Dual-Path Fusion references. Alternative pathways are compared separately. The future-force head in 2 provides training-only supervision. Force and torque components together form the six-axis force observation.
Method
Weight Sort
CoM Sort
Faucet Turn
Faucet Direction
Stove Turn
Stove Press
Drawer Pull
Nut Thread
Mean ↑
π0.5 (No Force) [ 3 ]
54.0
45.5
9.5
31.5
8.5
1.0
4.5
31.5
23.25
π0.5 ( + Token Fusion
50.5
50.0
12.0
28.5
6.0
2.5
13.0
26.0
23.56
π0.5 ( + Prompt Fusion
88.0
56.5
21.0
41.5
16.0
2.5
23.0
34.0
35.31
π0.5 ( + Cross-Attention
54.5
51.0
6.5
33.0
5.5
2.5
6.5
17.5
22.13
π0.5 ( + Token Fusion + Pred
78.0
48.5
13.5
32.5
6.0
1.0
4.0
26.5
26.25
π0.5 ( + Temporal Prompt
90.0
69.5
17.5
31.0
7.0
21.5
2.5
22.0
32.63
TABLE I: Simulation task success rates (%), with 200 episodes per task–method pair. Mean averages the eight tasks equally.
Fig. 4: Task outcomes and state diagnostics (200 episodes per task–method pair). Bars show success (green) and labeled failure categories (ochre/purple). In (a,b,d), filled/open markers report success rates for the labeled physical conditions. In (c,e,f,g,h), dots represent episode measurements colored by success (green) or failure (ochre). In (c,e,g), shaded target intervals map to [0,1] ; reaching these intervals alone does not establish full success. In (f), angles are normalized by the direction-specific resistance-peak angle. In (h), offset is measured during descent at a nut height of 35 mm on a symmetric-log axis; the ellipsis omits tick labels only. Count pairs report full-task successes / episodes reaching the target assembly state (centered, upright, and at the target height); full success additionally requires release and withdrawal.
Property inference
Resistance-dependent interaction
Assembly
Method
Weight Sort
Plate Retrieve
Stamp Press
Plug Insert
Block Insert
Mean ↑
π0.5 (No Force)
30.0
20.0
40.0
60.0
50.0
40.00
Token Fusion
45.0
40.0
50.0
60.0
10.0
41.00
Prompt Fusion
70.0
40.0
65.0
80.0
60.0
63.00
Cross-Attention
60.0
30.0
40.0
50.0
55.0
47.00
Token Fusion + Pred
45.0
60.0
60.0
45.0
35.0
49.00
TABLE II: Real-world task success rates (%), with 20 trials per task–method pair. The mean weights all five tasks equally; bold denotes the highest rate in each column.
Fig. 5: Real-world tasks. (a) Plug Insert, (b) Block Insert, (c) Plate Retrieve, (d) Stamp Press, and (e) Weight Sort. Each panel shows key interaction frames from the external camera.
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
Method
Property
Resistance
Assembly
Total
No Force
49.75
11.00
31.50
372
Token Fusion
50.25
12.40
26.00
377
Prompt Fusion
72.25
20.80
34.00
565
Cross-Attention
52.75
10.80
17.50
354
Token Fusion + Pred
63.25
11.40
26.50
420
Temporal Prompt
79.75
15.90
22.00
522
Appendix
TABLE III: Simulation category means (%) and total successes, derived from Table I . Each total is out of 1,600 episodes. Property, resistance, and assembly contain two, five, and one task, respectively.
Fig. 7: Six-axis wrench trajectories from three successful simulation demonstrations, each selected at seed 0 (episode 1). Top: Fx,Fy,Fz in newtons. Bottom: Tx,Ty,Tz in newton-metres. Gray intervals mark the logged lift-and-weigh phase in (a), resistance-rise phase in (b), and thread-engagement phase in (c). All 457, 530, and 917 control-rate samples are shown, respectively. These are illustrative expert trajectories, not comparisons between learned policies.
Fig. 8: Six-axis wrench trajectories covering all five real-world tasks. Each row shows one recorded demonstration: (a) Weight Sort (full can), (b) Plate Retrieve, (c) Stamp Press, (d) Plug Insert, and (e) Block Insert. Left: Fx,Fy,Fz in newtons. Right: Tx,Ty,Tz in newton-metres. All 424, 376, 675, 442, and 684 samples are retained, respectively, including offsets and transient variations. Time is relative to the first recorded sample; vertical scales vary by task and quantity. Colors and line styles follow Fig. 7 . The source records are episodes 0, 0, 1, 0, and 0 of the corresponding archived task collections, and are illustrative demonstrations rather than policy-evaluation rollouts.
Humans combine reasoning and motor control to solve complex manipulation tasks under diverse constraints. They build an understanding of the physical world that helps them convert reasoning into actions and quickly adapt to new scenes, tasks, and rules. We refer to this capability as intuitive manipulation. Existing benchmarks fail to capture this integration: they evaluate physical reasoning in isolation from execution, or measure policy performance without requiring explicit reasoning. We introduce IMBENCH, a benchmark designed to evaluate intuitive manipulation as an integrated capability spanning perception, physical reasoning, action generation, and iterative execution. Our tasks require models to infer task-relevant physical structure and generate feasible action sequences under explicit constraints, including contact-rich manipulation, tool use, and multi-stage dependencies. We introduce a benchmark of 35 tasks, 14K filtered trajectories, and scalable tools for generating diverse scenarios. Experiments reveal a consistent gap: vision language models show partial physical reasoning ability but fail to produce executable plans, while state-of-the-art vision-language-action models struggle to satisfy task constraints and generalize across scenarios. These results identify intuitive manipulation as a missing axis in current foundation models and generalist robot policies, and position IMBENCH as a step toward evaluating and enabling more integrated, adaptive physical 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.
Yang Li, Zhaxizhuoma, Hongru Jiang +11
Tongji University · Shanghai AI Laboratory · Lumos Robotics +6
Human demonstrations are a scalable data source for learning robot manipulation policies. However, common sources of human demonstration data, such as motion-capture trajectories and internet videos, capture mostly motion and appearance while missing the contact forces that are critical for force-sensitive manipulation. In this paper, we introduce ForceBand, a low-cost wrist-worn sEMG system that turns human muscle activity into force-enriched demonstrations. We first collect a 10-hour multimodal dataset containing egocentric video, sEMG, IMU, and fingertip force measurements across diverse actions and objects. Using this dataset, we pre-train an EMG2Force model that predicts per-finger forces from sEMG and IMU signals. After a short user-specific calibration, users can collect target-task demonstrations using only ForceBand and video; EMG2Force then labels these demonstrations with per-finger force traces, producing force-augmented demonstrations for robot policy learning. Experiments show that ForceBand recovers fine-grained fingertip interactions with over 50% lower force prediction error than vision-based baselines and achieves an 87% success rate on pick, squeeze, and place tasks that require object-specific force control across objects with diverse shapes, sizes, and weights. Project website: https://forceband-emg.github.io
Botao He, Zhi Wang, Linna Kuang +8
Amazon FAR · University of Maryland · Johns Hopkins University