SeonghuJeon/shortstop-vtp-small-sticky-ctx025-b64

Shortstop VTP Small — sticky pad, context 0.25 s, batch 64

Final 10-epoch / step 35,831 checkpoint. DINOv3-S encoder and native VTP; 1911 training episodes and 100 validation episodes; effective batch 64. Five causal frames from all three cameras, repeating the first frame when history is unavailable.

Rollout SR has not been measured for this checkpoint. The older VTP B16/6-epoch model achieved 3/40 final successes, but its weights were deleted. That score must not be attributed to this release. Final offline validation: joint MAE 0.01150 rad, tool-command MAE 0.5517 mm.

Task: catch a flying object with an adhesive pad and release it in the left bin. The wrist-tracking reset uses a fixed launcher. The instruction is:

Catch the incoming object with either adhesive pad, then release it into the left bin.

The benchmark's training/evaluation distributions are still provisional. These are experimental checkpoint exports, not an official benchmark release. No dataset, optimizer state, credentials, or private implementation source is included.

Runtime requirements

The matching Shortstop experimental adapter is required. The benchmark lives in worv-ai/worv PR #124, which requires repository access. That PR defines the task; it does not contain every experimental model/history change used here. Do not assume a stock OpenPI installation or PR #124 alone reproduces this checkpoint's inference. Use the matching experiment runtime when running the commands below from projects/shortstop.

The observation rate is 20 Hz; each call executes three consecutive 60 Hz absolute joint commands. Action layout: left six joints, left tool command, right six joints, right tool command. Joint targets are radians; the tool channel follows the YAM finger-position convention in metres. For adhesive pads, values below 0.0175 enable attachment and an open command releases it. No proprioceptive input is used by the policy. A state field may still be required by the adapter protocol.

Download and load

from huggingface_hub import snapshot_download
folder = snapshot_download("SeonghuJeon/shortstop-vtp-small-sticky-ctx025-b64")
# Requires the matching internal Pounce experiment implementation:
from pounce_1.modeling import Policy
policy = Policy.from_checkpoint(f"{folder}/policy.pt", "cuda")

In the matching runtime, set POUNCE_PYTHON to the Pounce environment, then:

python scripts/serve_policy.py --policy pounce --checkpoint /path/to/download/policy.pt   --execution-horizon 3 --port 18777

policy.pt contains architecture configuration, normalization buffers and the model state dictionary. The checkpoint retains two task-embedding slots, but only the adhesive-pad task was trained. The two-second future head is an auxiliary training target; execution uses the current 50 ms motor output. This derivative contains DINOv3 weights and is subject to the included DINOv3 license.

Loading and inference verified (2026-10-27)

Anonymous download of revision 7f046449f0decfe49b1e690a6d7f1923540bfd9f, weight/statistics hash verification, actual model loading, and 6 finite-output inference calls on A100 80GB passed. See inference_check.json for output shapes and package versions. Inputs were synthetic RGB. This does not establish rollout SR, websocket integration, or installation from a clean checkout. The matching private experiment runtime is still required. Runtime differences and corrected commands are recorded in PR #124.