Dongkkka/KPI_60k_original

GR00T N1.7 — checkpoint 60k

Inference upload package from pure_gr00t_objects47_b512_80k_20261105_run4/checkpoint-60000. Actual saved training step: 60,000. Training stopped after this checkpoint; 80,000 in original experiment settings was the initial target.

• Camera: cam_left_head.
• State/action: arm_left (8), arm_right (8), lift (1); 17 dimensions total.
• Action horizon: 16. Relative-action processing: disabled.
• Language: object name only, 47 training language groups.
• VLM frozen; percentile normalization disabled; training-time RTC disabled.
• Saved model weights: BF16, two Safetensors shards.
• Training batch: 512 per GPU, 4 GPUs, global batch 2048.

Use this package's config, processor, embodiment IDs and statistics together with the compatible GR00T N1.7 implementation. Do not substitute settings from the earlier three-camera Drill model. The VLM processor may also require the Cosmos-Reason2-2B assets identified by the configuration.

This folder contains model weights and inference configuration, plus original experiment metadata. Optimizer states, distributed training recovery files and RNG states remain in the original checkpoint and are intentionally excluded. This package alone is not a full training-resume checkpoint.

Packaging verification checks file hashes and Safetensors shard/index consistency; it does not constitute an inference or robot-execution evaluation.