Force2Act-Sim

Official version
Papers and technical reports

Force2Act is a contact-rich manipulation policy from MAGIC Lab at the National University of Singapore. It uses flow matching to plan pose and wrench sequences and a shared mixture of force experts for fine-grained contact control. Force2Act-Sim provides five task-specific policies for a simulated Flexiv Rizon 4s in Genesis.

The global policy plans in chunks. Shared MoFE adjusts contact force at each control step.

Papers and technical reports

Paper / report unconfirmed

Official introductory materials are recorded; no version-specific paper or technical report has been verified.

Last checked: 2026-11-22

How to access and use

Public code resources
Public access
Model files
Public access
Training data
Public access

Code and training / fine-tuning

Access
Public access

Verified on · 2026-11-22

Code license
MIT

Verified on · 2026-11-22

Model files

Model files
Block pivotingAccessWeights licenseCommercial use
Allowed under cited termsTerms

Verified on · 2026-11-22

Press-push boxAccessWeights licenseCommercial use
Allowed under cited termsTerms

Verified on · 2026-11-22

Bottle wipingAccessWeights licenseCommercial use
Allowed under cited termsTerms

Verified on · 2026-11-22

Round plug insertionAccessWeights licenseCommercial use
Allowed under cited termsTerms

Verified on · 2026-11-22

USB-A insertionAccessWeights licenseCommercial use
Allowed under cited termsTerms

Verified on · 2026-11-22

Shared Mixture of Force ExpertsAccessWeights licenseCommercial use
Allowed under cited termsTerms

Verified on · 2026-11-22

Release information

SourceVerified on · 2026-11-22

Checks before deployment

The published results are for the Genesis simulation. A separate real-robot checkpoint and real-robot success rates are unconfirmed. Available weights or code do not establish plug-and-play robot compatibility. Check the required configuration, data and interfaces in the source documentation.

Deployment and usage resources

Inference mode
Local inference with a 15 Hz control loop
Dataset format
LeRobot v3
Runtime and frameworks
PyTorch
Deployment requirements
NVIDIA GPU and EGL for benchmark camera rendering

Architecture and inputs / outputs

Model type
Action policy
Architecture
Transformer
Action generation
Flow matching
Input
Language / RGB images / Robot state
Output
TCP pose and wrench with hybrid motion–force control
Released
November 16, 2026

Scope and training

Code license

MIT

Embodiment scope

Single embodiment

Implementation source

Official

Record identity

Official version

Output

Actions

Robot form

Robotic arm

Task scope

Multiple tasks

Training data sources

290 Genesis demonstrations for the five task policies and shared MoFE

Training

Imitation learning

Evaluation results

No verified independent benchmark results are recorded for this version or the selected weights.

Publisher evaluation · Reported; not independently reproduced

Force2Act-Bench · Block pivoting ↗

Success rate: 93.3 %

Genesis / Flexiv Rizon 4s

Six masses from 0.4 to 0.9 kg, 25 trials each. Success requires the object's Z axis to tilt by at least 88° from its initial axis direction, using the smaller axis angle (0–90°).

Publisher evaluation · Reported; not independently reproduced

Force2Act-Bench · Press-push box ↗

Success rate: 100 %

Genesis / Flexiv Rizon 4s

Inclines of 0°, 2° and 4°, 10 trials each. Success requires at least 155 mm along the specified table-frame push axis. Contact by another arm link, tool contact more than 3 mm below the box top plane, or box tilt above 20° fails the trial.

Submit a source or correction