Core setup
- Choose modalities for a specific research hypothesis.
- Define clock alignment, coordinate calibration and stream association.
- Maintain data versions and common training and evaluation splits.
Solutions
Multimodal Research
Bring vision, tactile sensing, hand pose, FMG + IMU and robot state into a shared experiment. Align timing and data versions, then use controlled comparisons to test whether added signals improve manipulation.
Measure what each signal adds.
Multimodal learning, dexterous manipulation and joint data-model research
- Vision-tactile manipulation learning
- Human demonstration to robot action mapping
- Cross-user FMG, IMU and vision experiments
Optional modules
- Independent force or pose reference sensors.
- Robot logs, failure episodes and corrective demonstrations.
- Third-party model APIs, offline algorithms and private compute integration.
What you receive
Synchronization package
Device clocks, rates, coordinate frames, extrinsics, alignment method and measured errors.
Layered dataset
Raw observations, algorithm estimates, human annotations and robot action labels kept distinct.
Ablation studies
Compare vision-only, individual added signals and combined inputs on the same task and split.
Research report
Benefits, failure conditions and performance across participants, dates and scenes, with the next collection requirements.
Acceptance criteria
Check timing and completeness before offline metrics. Claims about better robot performance require controlled physical tests; gesture-recognition scores are not task success rates.
Configure your system
Capture / Specification-based configuration
Thor-Quad
Four-camera visual-inertial capture
- 30 FPS per camera
- 4 x 1080P
- IMU sampling
- 300 Hz
- Nominal system weight
- Approx. 500 g
Capture / Specification-based configuration
Thor-Hex
Six-camera visual-inertial capture
- 30 FPS per camera
- 6 x 1080P
- IMU sampling
- 300 Hz
- Nominal system weight
- Approx. 650 g
Capture / Specification-based configuration
UMI Gripper
Handheld manipulation demonstrations
- Global-shutter video
- 60 FPS
- IMU sampling
- 500 Hz
- Width and GT tactile output
- 100 Hz
Capture / Specification-based configuration
Tactile Data Glove
Contact sensing and electromagnetic tracking
- Standard tactile layout
- 360 points
- Nominal tactile sampling
- 200 Hz
- Nominal tracking output
- 180 Hz
Capture / Research validation
FMG + IMU
Forearm sensing for multimodal research
- Pressure patterns and motion
- FMG + IMU
- Left and right forearm capture
- Bilateral
- First-person task context
- Ego pairing
Robotics / Project configuration
Atlas D
Desktop bimanual learning platform
- Leader-follower setup
- 4 arms
- Reference single-arm payload
- 1.5 kg
- Reference single-arm reach
- 626 mm
FMG records pressure patterns associated with muscle deformation. It is not EMG or direct contact-force measurement. The benefit of an added modality is a hypothesis to test.
Build with Actrium
Start with your task.
Define the devices, data and evaluation your project needs.