---
license: other
tags:
- robotics
- franka
- real-robot
- imitation-learning
- action-prediction
- pytorch
- 3da
---

# Franka Real-Robot Final Checkpoints

This public repository contains the final compact PyTorch checkpoints from the
Franka real-robot finetuning runs found under `checkpoints/realrobot` on the
CVLAB1 workspace. The upload keeps one highest-step compact `.pt` file for each
`robot-*` run that still had a compact checkpoint on disk.

The runs are DA3/GAM real-robot finetunes over the Franka task mixture:
`franka_pickplace`, `franka_chocomilk`, `franka_kitchen`, and `franka_pot`.
Several variants are included, covering the original multitask recipe,
non-idle/OpenVLA-style sampling, an old-code resume comparison, and the latest
`openvla_noidle_resume_0036000` continuation.

## Checkpoints

| Run | Final step | Checkpoint | Size | Logged eval |
|---|---:|---|---:|---|
| `franka_multitask/robot-20260525-061300` | `0020000` | [`.pt`](./checkpoints/franka_multitask/robot-20260525-061300/0020000.pt) | 3.0 GiB | l1_raw=0.0019, r2=0.180, rel@5/10=0.386/0.557 |
| `franka_multitask/robot-20260526-035255` | `0030000` | [`.pt`](./checkpoints/franka_multitask/robot-20260526-035255/0030000.pt) | 3.0 GiB | l1_raw=0.0069, r2=0.144, rel@5/10=0.330/0.498 |
| `franka_multitask/robot-20260526-234418` | `0034000` | [`.pt`](./checkpoints/franka_multitask/robot-20260526-234418/0034000.pt) | 3.0 GiB | l1_raw=0.0041, r2=0.192, rel@5/10=0.334/0.508 |
| `franka_multitask_0526/robot-20260527-064944` | `0056000` | [`.pt`](./checkpoints/franka_multitask_0526/robot-20260527-064944/0056000.pt) | 3.0 GiB | l1_raw=0.0021, r2=0.067, rel@5/10=0.401/0.534 |
| `franka_multitask_0526/robot-20260528-104747` | `0002000` | [`.pt`](./checkpoints/franka_multitask_0526/robot-20260528-104747/0002000.pt) | 3.0 GiB | l1_raw=0.0036, r2=0.064, rel@5/10=0.404/0.548 |
| `franka_multitask_oldcode_resume20k/robot-20260528-130248` | `0032000` | [`.pt`](./checkpoints/franka_multitask_oldcode_resume20k/robot-20260528-130248/0032000.pt) | 3.0 GiB | l1_raw=0.0019, r2=0.185, rel@5/10=0.372/0.553 |
| `franka_multitask_openvla_noidle_resume_0036000_8gpu_gb512_cvlab1/robot-20260529-151455` | `0052000` | [`.pt`](./checkpoints/franka_multitask_openvla_noidle_resume_0036000_8gpu_gb512_cvlab1/robot-20260529-151455/0052000.pt) | 3.0 GiB | l1_raw=0.003, r2=0.250, rel@5/10=0.400/0.574 |

## Included Files

- `checkpoints/<experiment>/<robot-run>/<step>.pt`: compact finetune checkpoint
  intended for inference/deployment style loading.
- `checkpoints/<experiment>/<robot-run>/config.yaml`: resolved trainer config
  saved with the run.
- `checkpoints/<experiment>/<robot-run>/log.txt`: local training log for that
  run.
- `checkpoints/<experiment>/<robot-run>/metadata.json`: generated upload
  metadata, file size, source path, and parsed eval summary where available.
- `configs/training/realrobot/`: real-robot training YAMLs and DeepSpeed config
  copied from `GAM_realrobot_train`.
- `stats_realrobot/`: action/proprio stats and non-idle range metadata copied
  from `GAM_realrobot_train/_stats_realrobot`.
- `manifest.json`: machine-readable index for this upload.

## Notes

The DeepSpeed ZeRO optimizer/model-state directories are not uploaded. They are
available locally next to each `.pt` checkpoint, but this Hub repo is scoped to
the compact final checkpoints plus the configs needed to identify and deploy
them.

Some saved configs contain absolute paths from the original training machine.
Treat those paths as provenance; the companion stats/config folders in this repo
provide the portable context.

Repository ID: `SeonghuJeon/franka-realrobot-final-checkpoints`