UMI-Benchmark-v1: 20,000 Real-World Robot Manipulation Sessions Across 10 Tasks

UMI-Benchmark-v1 contains 20,000 real-world robot manipulation sessions. The dataset covers 10 distinct tasks, split into 4 single-arm and 6 dual-arm activities, with sessions ranging from 1,991 to 3,000 per task. It was created by UMIbenchmark and last updated on June 10, 2026.

Use Cases

• Benchmarking single-arm robot manipulation performance based on the 4 single-arm tasks described
• Training models for dual-arm coordination based on the 6 dual-arm tasks described
• Analyzing task performance across a large-scale, real-world session corpus of 20,000 sessions
• Developing generalizable robot policies using the multi-task structure of the dataset

Strengths

• Contains 20,000 real-world robot manipulation sessions, providing substantial scale
• Covers 10 distinct tasks, offering variety in manipulation challenges
• Includes a balanced split of 9,988 single-arm and 10,012 dual-arm sessions

Limitations

• Column-level documentation is absent; field semantics must be inferred after download
• Row count per individual task beyond the provided examples is unknown, which may limit suitability assessment

Provenance

Source

huggingface

Freshness

Last updated 2026-06-10 03:44:06; freshness should be verified