UR7e Sort RGBBlock DCT-tuned SmolVLA 3000-step Checkpoint
This repository preserves the DCT-tuned SmolVLA checkpoint and Method3 Phase2 server artifacts for the UR7e Sort RGBBlock task.
Source
- Phase1 dataset:
CoRL2026-CSI/UR7e_phase1_Sort_RGBblock_10fps - Task: Sort each colored block onto the dish of the matching color.
- Training:
steps=3000,batch_size=16,gradient_accumulation_steps=1,empty_cameras=1 - Camera rename:
realsense_wrist -> camera1,realsense_topview -> camera2
Contents
pretrained_model/: LeRobot/SmolVLAfrom_pretrainedcheckpoint.skill_dct/: raw and canonical skill DCT sidecar parquet files.vector_db/: Phase2 P_phase1 vector DB and server buffer DB.session/: minimal Phase1 session artifacts needed by UR7e Phase2 replay.config/: Phase2 and recording configs used for server startup.MANIFEST.json: file hashes, training metadata, and DB summary.
Verified DB Summary
- Skills:
29 - Entries:
580 - Per-skill entries:
20to20 - Key shape example:
[20, 966]
Server Startup Hint
cd /home/csiwoo/CORL2026/ur7e_autodata_workspace/runtime/AutoDataCollector-ur7e
PATH=/home/csiwoo/robobridge/nvidia_580_126_09/usr/bin:$PATH \
LD_LIBRARY_PATH=/home/csiwoo/robobridge/nvidia_580_126_09/usr/lib/x86_64-linux-gnu:${LD_LIBRARY_PATH:-} \
SKIP_SETUP=1 GPU_ID=0 ENV_NAME=gpu_server HOST=0.0.0.0 PORT=50061 \
RECORDING_CONFIG=pipeline_config/recording_config_ur7e_phase2.yaml \
URDF=assets/urdf/ur7e.urdf \
bash grpc_server/run_server.sh