Prompt Fusion · simulation
Highest observed mean across eight tasks; no-force baseline: 23.25%.
September 2026
How should a policy use force?
Results from ForceBench: Benchmarking Force Integration for Robotic Manipulation. Seven configurations share a π₀.₅ backbone to compare a no-force baseline with six force integration designs. Additional reference systems are reported separately.
The shared-backbone comparison includes changes in representation, normalization, capacity, and insertion location. Temporal designs serve as separate reference points; the predictive variant retains the Token Fusion input pathway.
Prompt / Dual-Path · real world
Tied highest observed mean across five tasks; no-force baseline: 40.00%.
Benefits depend on the task
Physical-property inference, resistance-dependent interaction, and contact-rich assembly.
Table I · 16,000 evaluation episodes
Simulation
10 systems · 8 tasks
200 episodes per task–method pair
Bold marks the best observed rate among the seven π₀.₅ configurations. Additional references stay in a separate block; π₀ reaches 36.0% on Nut Thread, versus 34.0% for the best π₀.₅ configuration. Click a column heading to sort within each block.
| Shared-backbone comparison · π₀.₅ | |||||||||
|---|---|---|---|---|---|---|---|---|---|
| π₀.₅ (No Force) | 54.0 | 45.5 | 9.5 | 31.5 | 8.5 | 1.0 | 4.5 | 31.5 | 23.25 |
| Token Fusion | 50.5 | 50.0 | 12.0 | 28.5 | 6.0 | 2.5 | 13.0 | 26.0 | 23.56 |
| Prompt Fusion | 88.0 | 56.5 | 21.0 | 41.5 | 16.0 | 2.5 | 23.0 | 34.0 | 35.31 |
| Cross-Attention | 54.5 | 51.0 | 6.5 | 33.0 | 5.5 | 2.5 | 6.5 | 17.5 | 22.13 |
| Token Fusion + Pred | 78.0 | 48.5 | 13.5 | 32.5 | 6.0 | 1.0 | 4.0 | 26.5 | 26.25 |
| Temporal Prompt | 90.0 | 69.5 | 17.5 | 31.0 | 7.0 | 21.5 | 2.5 | 22.0 | 32.63 |
| Dual-Path Fusion | 99.5 | 72.0 | 22.0 | 30.0 | 17.5 | 19.0 | 4.5 | 16.5 | 35.13 |
| Additional reference systems | |||||||||
| π₀ | 61.5 | 39.0 | 3.5 | 27.0 | 7.0 | 5.0 | 7.5 | 36.0 | 23.31 |
| ForceVLA | 35.5 | 45.5 | 12.5 | 14.0 | 10.0 | 12.0 | 20.0 | 26.0 | 21.94 |
| ForceFlow | 51.5 | 3.0 | 0.0 | 0.5 | 0.0 | 0.0 | 0.0 | 0.0 | 6.88 |
Rows follow the manuscript order.
Simulation evaluation protocol
A separate policy is trained for every task. The seven π₀.₅ configurations share task demonstrations and are trained for 10,000 steps with a batch size of 32. Training instances use seeds 0–199; evaluation uses the same seeds 300–499 for every method.
Physics runs at 60 Hz and control at 15 Hz, with at most 1,500 control steps (100 simulated seconds). The π₀.₅ configurations, π₀, and ForceVLA predict and execute 50-step action chunks. ForceFlow predicts 64 steps and executes 32 before the next query. The control rate is not the policy-query rate.
Additional reference systems retain differences in training and execution settings. All ten systems use the same lower-bound-only angle criterion for Faucet Direction.
Table II · 800 evaluation trials
Real-world manipulation
8 systems · 5 tasks
20 trials per task–method pair
Bold marks the highest observed rate in each column, including ties. Prompt Fusion improves success over the no-force baseline on all five tasks. Both temporal designs improve Weight Sort but underperform the baseline on Block Insert.
| Shared-backbone comparison · π₀.₅ | ||||||
|---|---|---|---|---|---|---|
| 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 |
| Temporal Prompt | 80.0 | 45.0 | 65.0 | 75.0 | 35.0 | 60.00 |
| Dual-Path Fusion | 95.0 | 45.0 | 70.0 | 85.0 | 20.0 | 63.00 |
| Additional reference system | ||||||
| ForceVLA | 50.0 | 0.0 | 40.0 | 70.0 | 5.0 | 33.00 |
Rows follow the manuscript order.
Real-world evaluation protocol
Experiments use a Franka Research 3 with a six-axis wrist force/torque sensor. Each system uses 200 demonstrations per task and is trained for 10,000 steps with a batch size of 64. All systems are evaluated on the same 20 initial conditions per task.
Reading the results
Observed performance
These are the point estimates reported in the submitted manuscript. A small difference in mean success, such as 35.31% versus 35.13%, does not by itself establish statistical superiority. The tables do not report confidence intervals or variation across training seeds.
Complete-task success
Reaching a target object state does not always complete a task: release, withdrawal, and other task-specific requirements can still fail. The manuscript pairs success rates with outcome and trajectory diagnostics.
Source: Tables I and II of the manuscript submitted on September 15, 2026. These are manuscript results, not entries from an open community leaderboard. Values are also available in the structured results file.