---
license: apache-2.0
tags:
  - robotics
  - vision-language-action
  - robosynchallenge
  - turbovla
  - diffusion-policy
---

# RoboSynChallenge preliminary submission

Task-specific checkpoints for team `scv-think-tank` (`HanJW2000`).  The
submission uses six TurboVLA checkpoints and one Diffusion Policy checkpoint
trained only with the released RoboSynChallenge simulation data plus generated
hard-position trajectories for `click_bell`.

## Task mapping

| Task | Policy | Checkpoint | Local seed-0 result |
|---|---|---|---:|
| `table_rearrangement` | TurboVLA | `checkpoints/table_rearrangement/turbovla/final_model/ema_ema_pytorch_model.pt` | 99/100 |
| `handle_basket` | TurboVLA | `checkpoints/handle_basket/turbovla/final_model/ema_ema_pytorch_model.pt` | 96/100 |
| `mixer_operating` | TurboVLA | `checkpoints/mixer_operating/turbovla/final_model/ema_ema_pytorch_model.pt` | 95/100 |
| `item_assembly` | TurboVLA | `checkpoints/item_assembly/turbovla/final_model/ema_ema_pytorch_model.pt` | 76/100 |
| `water_pouring` | TurboVLA | `checkpoints/water_pouring/turbovla/final_model/ema_ema_pytorch_model.pt` | 74/100 |
| `items_handover` | TurboVLA | `checkpoints/items_handover/turbovla/final_model/ema_ema_pytorch_model.pt` | 64/100 |
| `click_bell` | Diffusion Policy | `checkpoints/click_bell/dp/pretrained_model/` | 66/100 |
| `manipulate_pipette` | official ACT fallback | `RoboSynChallenge/ACT_sim_manipulate_pipette` at `ca013b2e202d594bf347530c61c3ba28c06d1bc2` | official 61/100 |
| `drawer_open_place` | official SmolVLA fallback | `RoboSynChallenge/SmolVLA_sim_drawer_open_place` at `c0088d84a568f93fb4401aabafcc41cf643efcdd` | official 62/100 |
| `sample_loading` | official SmolVLA fallback | `RoboSynChallenge/SmolVLA_sim_sample_loading` at `2a4f41ab3be8e5f544c890d6054bc8de3be5c53b` | official 6/100 |

The reported aggregate of the selected public results is 69.9% success.  The
organizer's held-out preliminary evaluation may differ from these public-seed
results.

## Shared TurboVLA dependencies

The upload includes the exact local model files used during evaluation:

- `dependencies/dinov3-vitl16/`
- `dependencies/bert-base-uncased/`
- `dependencies/groundingdino/`

The code repository README documents the environment and complete evaluation
commands.  `task_checkpoints.json` is the machine-readable task router.

## Data

- Official released RoboSynChallenge simulation datasets for all tasks.
- An additional 450 `click_bell` hard-position trajectories generated with the
  official simulator and task assets.
- No external demonstration dataset.

## Integrity

`SHA256SUMS` records every uploaded checkpoint and dependency model file.