DozenDucc — Waddle demos were pure code-as-policy, no robotics models under the hood

TL;DR

Stone Tao (@DozenDucc, Waddle Labs) confirms in a reply thread that every demo in the Waddle announcement was code-as-policy — the LLM agent writes control programs and no separate robotics model (VLA / world-action model) was loaded to execute them. He acknowledges that extreme-diversity tasks like crumpled-shirt folding are genuinely hard for code-as-policy, but describes a candidate approach (programmatically grab corners, then whip the shirt in the air to flatten) and flags the open problem as task-completion detection — knowing when the shirt is actually flat.

Key claims

“All the demos here were code-as-policy. No robotics models.” [tweet 2/3]
Extreme-diversity manipulation (e.g. crumpled shirts) is tricky with pure code, but tractable via programmatically-composed primitives like grab-corners + whip-to-flatten [tweet 2/3].
The unresolved bottleneck for such recipes is not motion generation but state detection — how the agent knows the shirt is flat [tweet 3/3].

Method

Reply thread on X to @Stone_Tao’s skeptical question about whether the Waddle demos were “just loading an existing robotics model.” DozenDucc’s response is a two-tweet clarification, no code or media. No embedded artifact URLs.

Results

None reported — this is a positional clarification, not a result. The concrete evidence being clarified is the demo suite in Introducing Waddle: Agents that Control Robots (t-shirt folding, LEGO pick-and-place ×1000 overnight, microswitch placement).

Why it’s interesting

This tweet is a load-bearing clarification on top of Introducing Waddle: Agents that Control Robots — the Waddle post claims agents can call VLAs as tools, but the shipped demos apparently didn’t use one at all. That sharpens Waddle’s positional stance against end-to-end VLA Models considerably: it is not “agent + VLA is better than VLA alone” but “the agent alone, with good primitives, was enough for the demos we chose.” It also names the open frontier for the code-as-policy revival explicitly — perception of task completion — which is the exact gap VIA: Visual Interface Agent for Robot Control (VIA) and CaP-X: A Framework for Benchmarking and Improving Coding Agents for Robot Manipulation (CaP-X) are trying to close from the interface / benchmark side. Dejia’s bufo-annoyed emoji is a fair reaction: the “no robotics model” admission makes the demos look further from a general robot policy than the launch post implied.