Skip to main content

MiniCPM-RobotTrack

MiniCPM-Robot Embodied Series

MiniCPM-RobotTrack uses EVT-Bench to evaluate static target tracking (STT), dynamic target tracking (DT) and adversarial target tracking (AT). Metrics include success rate (SR), tracking rate (TR) and collision rate (CR).

Metrics​

MetricMeaningDirection
SRSuccess rateHigher is better
TRTracking rateHigher is better
CRCollision rateLower is better

Model and Vision Dependencies​

Place the full model snapshot in minicpm_robot_track/checkpoints/MiniCPM-RobotTrack/. Evaluation also requires DINOv3 ViT-S/16 and SigLIP So400m; you can point to local directories via environment variables.

Bash
export DINOV3_MODEL_PATH=/path/to/dinov3-vits16
export SIGLIP_MODEL_PATH=/path/to/siglip-so400m-patch14-384

Running the Evaluation​

Bash
cd MiniCPM-RobotTrack
CKPT=minicpm_robot_track/checkpoints/MiniCPM-RobotTrack
bash scripts/eval_stt.sh "$CKPT" results/stt
bash scripts/eval_dt.sh "$CKPT" results/dt
bash scripts/eval_at.sh "$CKPT" results/at

Official Results Summary​

TaskSRTRCR
STT84.189.83.0
DT53.273.413.6
AT58.080.49.0

The values in the table are from results published in the official repository, in percent. When reproducing, record the scene assets, software versions, random seeds, model snapshot and metric implementation.

Reproduction Checklist​

  • Habitat-Sim, Bullet, PyTorch and CUDA versions recorded

  • HM3D, MP3D, humanoid and robot asset directories complete

  • DINOv3, SigLIP and MiniCPM-RobotTrack weight versions pinned

  • STT, DT and AT output directories kept separate

  • SR, TR and CR computed the same way as the official implementation

TopicLink
Official evaluation codeOfficial evaluation code
EVT-Bench results tableEVT-Bench results table