Feelsight: Visuo-Tactile Robot Manipulation Data from In-Hand Rotation

Feelsight is a dataset of vision, touch, and proprioception data collected from in-hand rotation of objects via an RL policy. It consists of 70 experiments, with 30 in the real-world and 40 in simulation, each lasting 30 seconds. The dataset was created by author 'suddhu' and was last updated on Hugging Face in November 2024.

Use Cases

• Training neural field models based on the multimodal sensor data mentioned in the description
• Developing RL policies for in-hand object rotation based on the stated collection method
• Benchmarking sim-to-real transfer for robotic manipulation using the 30 real-world and 40 simulation experiments
• Researching visuo-tactile fusion for robotic perception and control

Strengths

• Contains 70 total experiments, providing a defined scale for research
• Includes data from both real-world (30 experiments) and simulation (40 experiments) environments
• Each experiment provides a consistent 30-second time series of sensor data

Limitations

• Column-level documentation is absent; field semantics must be inferred after download
• Row count is unknown, which may limit suitability assessment
• Description metadata is limited; actual data quality requires manual inspection after download

Provenance

Source

suddhu on Hugging Face

Collection Method

Collected from in-hand rotation of objects via an RL policy

Freshness

Last updated 2024-11-11 04:41:26