MiniCPM-RobotManip
MiniCPM-Robot Embodied Series
MiniCPM-RobotManip provides reproducible evaluation integrations for four benchmarks: LIBERO, CALVIN, RoboTwin2 and RMBench. The model server and each simulator run separately, exchanging observations, instructions and actions over WebSocket and MessagePack.
Supported Benchmarks
| Benchmark | Purpose |
|---|---|
| LIBERO | Long-horizon, multi-task robot manipulation evaluation |
| CALVIN | Language-conditioned continuous manipulation task evaluation |
| RoboTwin2 | Dual-arm manipulation and randomized scene evaluation |
| RMBench | Comprehensive robot manipulation evaluation |
Reproduction Workflow
- Install the MiniCPM-RobotManip model environment and prepare the weights.
- Install the simulator, datasets and assets following the official instructions for the target benchmark.
- Start the MiniCPM-RobotManip model server.
- Start the corresponding evaluator in a separate environment and connect it to the model server.
- Record software versions, random seeds, task splits, success-rate definitions and logs.
Simulators, datasets and scene assets are external dependencies. When reproducing results, pin dependency versions and follow the license and data-use terms of each project.
Official Results Summary
| Model | LIBERO | CALVIN ABC→D | RMBench |
|---|---|---|---|
| MiniCPM-RobotManip 1.5B | 97.5 | 4.1 | 53.3 |
The values in the table are from results published in the official repository. Before comparing across models, confirm that evaluation settings, task splits and metric definitions are consistent.
Related resources
| Topic | Link |
|---|---|
| Evaluation overview | Evaluation overview |
| Official results table | Official results table |