DAVATaR: Depth-Aware Visually Augmented Tactile Learning Across the Air-Water Interface¶
Session: Not recorded. A secondary listing (hongsong-wang.github.io/CoRL2026) marks it as a Spotlight (unverified). Speaker(s): Paper: PDF on the project page (anonymized submission version, 16 pages) by Bowen Jiang*, Siqi Shang*, Rahul Ajay Iyer, Ashwin Prakash, Lillian Chin†, and Roberto Martín-Martín† (UT Austin; * equal contribution, † equal advising). The author list comes from the project page. No arXiv ID yet. The same secondary listing gives OpenReview hSRtOgnkTM (not opened). Project page: davatar-corl26.github.io, with a 3 min overview video. Code and data are not mentioned.
Paper summary (from the PDF text; figures not inspected)¶
- Problem: Submerged manipulation means a robot in air reaches into water, for example washing dishes, bathing, or water sampling. This differs from fully underwater robots because the robot crosses the air–water interface over and over.
- Vision breaks: Refraction shifts and rescales objects, reflection causes glare, and foam, bubbles, or turbidity occlude the scene.
- Electronics and water don't mix: Conventional tactile sensors put electronics in the wet region.
Method
- Gripper:
- Each finger is a 305 mm aluminum H-beam with a 90 mm compliant fingertip printed in EPU40, so the electronics sit far from the water.
- Contact force: Uses fluidic innervation. Each fingertip has 5 pairs of embedded L-shaped air channels, each pair read by a differential pressure transducer (DLHR-L05D) at the finger root. The differential layout cancels temperature and hydrostatic offsets.
- Immersion depth: A tube open to the water and a sealed reference tube run along the back of the finger and feed one differential transducer. Hydrostatic pressure gives depth. They turn the "nuisance" of hydrostatic pressure into a signal.
- In total the gripper gives 11 signals at 20 Hz: 10 force and 1 depth.
- Submerged-domain augmentations (wrist RGB only, training time only, on top of translation and color jitter):
- Reflection-like: Translucent bright patches covering about 6% of the image, with random brightness and scattering. Applied with p = 0.3.
- Refraction-like: Random zoom in U(0.95, 1.05), cropped or padded back to size. Applied with p = 0.3.
- Policy: Behavior-cloning transformer with an MDN head (5 components). It predicts operational-space end-effector deltas plus a gripper command.
- Inputs are a history of H = 5 steps. The 256×256 wrist image goes through ResNet-18 with spatial softmax (32 keypoints) to make one token, and the 11-d tactile vector goes through an MLP to make one token. A learned readout token fuses them.
- It is small: 64-d tokens, 4 layers, and modality dropout of 0.2 on tactile.
- Trained from 40 SpaceMouse demos per task on a Franka Panda. Evaluation runs at 5 Hz.
Results
- Sensor characterization:
- Force RMSE overall is 0.37 N for grip and 0.66 N for push, over a 0–20 N range. Error grows with depth, from 0.11 N for grip in dry conditions to 0.85 N at 20 cm.
- Depth is near-linear (R² = 0.9996), with 1.6 mm RMSE.
- Policies: Four tasks with 18 rollouts per method and condition (3 seeds × 6). Demos come only from the normal setting, and the adversarial settings are held out. Each entry is the mean of the normal and adversarial results.
| Method | Water sampling | Spoon pickup | Sink plug removal | Plate wiping | Mean |
|---|---|---|---|---|---|
| VisionOnly | 0.11 | 0.40 | 0.24 | 0.37 | 0.28 |
| DAVATaR-T (no tactile) | 0.22 | 0.52 | 0.41 | 0.48 | 0.41 |
| DAVATaR-VA (no augmentation) | 0.56 | 0.61 | 0.37 | 0.64 | 0.55 |
| DAVATaR | 0.67 | 0.75 | 0.61 | 0.81 | 0.71 |
- Adversarial conditions are where it matters: Success averaged over the adversarial settings is 9% for VisionOnly, 25% without tactile, 37% without augmentation, and 61% for the full method. Sink plug removal under adversarial conditions goes from 0% (VisionOnly) to 50% (full).
- Why tactile helps:
- In water sampling, foam hides the surface, so vision can't tell how deep the syringe is.
- In spoon pickup, refraction makes the gripper look closer than it is, so vision-only policies close early.
- Attention rollout shows tactile attention rising during contact phases. The authors call this qualitative only.
- Single-augmentation ablation (spoon pickup):
- Reflection augmentation alone fixes glare: without tactile it goes from 0% to 54%, and with tactile from 46% to 71%.
- Refraction augmentation alone barely helps without tactile (38% to 39%), but helps a lot with tactile (37% to 67%).
- The authors' reading: reflection hurts visibility, while refraction corrupts geometry, which touch can fix.
- Limitations (as the authors state them):
- The compliant fingertips deform, and the channel sealant peels under repeated large forces.
- The long fluidic lines drift with temperature (ideal gas law).
- The augmentations model only zoom and patches. They miss water motion, dynamic lensing, and defocus.
Talk notes¶
- The talk opened with the paper's Fig. 1: "Submerged tasks are difficult." Submerged manipulation suffers air–water interface distortions: refraction, reflection, and foam.