README.md

metadata

title: RynnValue 4B
emoji: 🤖
colorFrom: blue
colorTo: indigo
sdk: gradio
sdk_version: 5.50.0
app_file: app.py
pinned: false
license: apache-2.0
short_description: How far is the robot from finishing the task?
python_version: '3.12'
startup_duration_timeout: 1h
models:
  - Alibaba-DAMO-Academy/RynnValue-4B
tags:
  - robotics
  - value-model
  - video

RynnValue-4B

Interactive demo of Alibaba-DAMO-Academy/RynnValue-4B, a general-purpose value model for robot manipulation.

Feed it a manipulation video plus the task instruction and it predicts, along prefixes of the video, how many seconds of work are still left until the instruction is complete, and generates a short analysis block (video description, does the video match the instruction, did the agent succeed).

The inference protocol mirrors rynn_infer/inference.py from the official repo: prefix-uniform sampling — each score conditions only on frames seen so far — followed by a generate() pass over the full-video prefix for the analysis block. The custom pred_slot_isolated_eager attention required by the value heads is forced explicitly, as in the reference script.

Example assets

• examples/put_box_in_drawer.mp4 — transcoded (854×480) copy of example/Put_the_box_in_the_drawer_and_close_it.mp4 from the RynnValue repository (Apache-2.0).
• examples/soar_put_green_stick_in_brown_bowl.mp4, examples/berkeley_rpt_stack_cup.mp4, examples/jaco_play_pick_up_green_cup.mp4 — from the Robometer example videos bundled in the same repository (robometer/, MIT License), originally derived from Open X-Embodiment trajectories.