RoboCurve lists Waddle-1 at 81.4% mean progress on RoboDojo-RC Tier 1, ahead of GPT-6.2 Astra at 65.7%, per robocurve.org on November 20, 2026. Waddle Labs promoted 81 in posts from @thewaddlelabs the same day. The score is on the official board with trial artifacts; the comparison is still not apples-to-apples because RoboCurve runs Waddle-1 without the 40-call cap it enforces on frontier LLM rows.
TL;DR
| Question | Answer |
|---|---|
| Waddle Labs social claim | Waddle-1 at 81 on RoboDojo-RC Tier 1 |
| RoboCurve official row | Waddle-1 at 81.4% mean progress with logs on robocurve.org |
| Previous top frontier LLM row | GPT-6.2 Astra 65.7% under 40-call cap |
| Cap difference | Waddle-1: one chat per task, 10-minute limit, no model-call cap |
| Benchmark | Six real bimanual tasks adapted from RoboDojo |
| Product status | Preview API per Waddle Labs posts |
| Compare to uni-robot-agent? | Different harness and task slice; see our earlier post |
What Waddle Labs says Waddle-1 is
Aggregator summaries attribute these details to @thewaddlelabs and @robocurve posts on November 20, 2026:
- Waddle-1 controls bimanual robots on unseen tasks zero-shot.
- It uses multiple cloud LLMs with one shared context across trials (per @robocurve's thread description).
- Lower-level policies run locally while the LLM directs tool use.
- A preview API is open to developers.
Per Waddle Labs posts, product details still live on X and RoboCurve's leaderboard row; there is no stable public docs site linked from those sources yet.
What RoboCurve's leaderboard shows today
RoboCurve maintains RoboDojo-RC Tier 1, an adaptation of six real RoboDojo manipulation tasks on I2RT YAM arms. Its November 20, 2026 table includes:
| Model | Mean score (RoboCurve) | Eval notes |
|---|---|---|
| Waddle-1 | 81.4% | Full system; no model-call cap; one chat per task |
| GPT-6.2 Astra | 65.7% | 40 model calls per attempt cap |
| Dream-1 | 59.5% | Hosted VLA stack |
| Claude Fable 5.2 | 57.5% | 40 model calls per attempt cap |
| Gemini 4 Carbon | 54.5% | 40 model calls per attempt cap |
Each entry includes per-trial cost, time, logs and video on RoboCurve's public runs page. RoboCurve's site labels Waddle-1 explicitly as a full system that controls the arm through lower-level policies with one shared context per task, a 10-minute wall clock, and no call cap, unlike the 40-call limit on frontier LLM configurations.
The social 81 rounds the published 81.4% row. The verified question is not whether the number exists, but whether you should rank it beside capped LLM scores without reading the footnotes.
Why 81.4% vs 65.7% is not a simple "beats GPT" headline
Four nuances matter for practitioners:
-
Subset benchmark — RoboDojo-RC is six tasks, not the full 18-task real RoboDojo suite or 42-task simulation suite. RoboCurve documentation stresses scores are not directly comparable to the original RoboDojo paper tables.
-
Unequal caps — RoboCurve caps frontier LLM configs at 40 model calls per attempt while Waddle-1 runs with no call cap under a 10-minute limit. That is disclosed on the leaderboard, but headlines often skip it.
-
Harness effects — Our uni-robot-agent post showed the same frontier model can jump from 28.97 to 51.90 on simulation RoboDojo when the harness changes. Waddle-1 is a vendor full stack (cloud LLM plus local policies), not the same Inspect Robots LLM-only harness used for Astra's row.
-
Metric definition — RoboCurve reports mean rubric progress across 120 task attempts per configuration on Tier 1. Waddle's published row uses the same attempt count; the rubric matches other Tier 1 entries on RoboCurve's site.
Robotics gentle grip illustrating bimanual manipulation agents
How this fits the embodied-AI race
Frontier labs treat real-robot tool-calling as the next leg after simulation leaderboards. AA-Robotics v0.1 already showed humanoid scores collapsing versus arm tasks. RoboDojo-RC focuses on reachable bimanual chores — cap a pen, sort objects, stack bowls — where LLM agents with enough steps can look competent.
With an official 81.4% row, Waddle-1 is another startup full stack ahead of published frontier LLM scores on real hardware, similar in narrative to uni-robot-agent's Qwen headline but on RoboCurve's RC Tier 1 slice — still subject to the cap and harness caveats above.
Meta's unrelated consumer app Waddle World in the Philippines is a different product; do not conflate it with Waddle Labs' robot agent name collision.
The six RoboDojo-RC Tier 1 tasks
RoboCurve adapted these real-robot chores from the broader RoboDojo suite (see the RoboDojo paper on arXiv 2607.04434 for the full 42 simulation plus 18 real tasks):
| Task name (RC) | What it exercises |
|---|---|
| Cap Pen | Precision grasp and placement on a small object |
| Pack And Pour Fruit | Pouring and packing with fluid-like objects |
| Sort Objects | Classification and bin placement |
| Stack Bowls | Stacked stability with repeated picks |
| Stand Up Bottles | Uprighting fallen cylinders |
| Store In Safe | Longer horizon storage inside a container |
RoboCurve uses I2RT YAM bimanual arms and Inspect Robots as the execution framework. Progress is scored with staged rubrics rather than a single success bit, which is why mean percentages like 65.7% still represent partial competence.
When Waddle Labs quotes 81, ask whether that is the same mean rubric progress across the same 120 attempts cap RoboCurve publishes for frontier models.
Leaderboard transparency: what RoboCurve gives you
Unlike simulation-only boards, RoboCurve's public site emphasizes per-trial artifacts:
- Wall-clock time per task attempt.
- Estimated API cost for LLM-driven agents.
- Logs and video reviewers can replay.
That is why explainx.ai treats an unpublished X score as lower trust than an official row. RoboCurve is effectively a registrar for embodied evals; vendors market numbers, but the board curates evidence.
Agent architecture questions for Waddle-1
Until Waddle Labs ships detailed docs, these are the questions practitioners are asking (paraphrased from launch discussion, not verified specs):
- Which LLM APIs can sit in the planner slot, and is model switching allowed mid-episode?
- How often does the cloud planner call local policies — every control step or at phase boundaries?
- What safety supervisor enforces joint limits and e-stop on hardware?
- Does the preview API expose only simulation, or real robot tenants with liability terms?
- How is context shared across trials — full chat history or summarized state?
Answers change whether Waddle-1 is comparable to uni-robot-agent tool routing or to monolithic vision-language-action models like Dream-1 on the same arm platform.
Simulation vs RC Tier 1 vs full real RoboDojo
| Suite | Setting | Task count | Notes |
|---|---|---|---|
| RoboDojo simulation | Isaac Sim | 42 | Physics pauses while the model thinks |
| RoboDojo real (original) | ARX, Piper arms | 18 | Full paper protocol |
| RoboDojo-RC Tier 1 | I2RT YAM | 6 adapted | RoboCurve rubrics; not comparable to 18-task scores |
Our uni-robot-agent article covers simulation RoboDojo with a different harness. Do not chain those percentages to Waddle's 81 without reading RoboCurve's comparability footnotes.
Policy and safety angle
Real-robot APIs that accept cloud LLM plans need rate limits, workspace bounds, and human oversight no matter how high the benchmark score is. Teams shipping warehouse pilots should pair vendor claims with on-site e-stop testing, the same way Italy's Tesla assisted-driving approval process treats autonomy as a regulatory object, not a leaderboard badge.
What people are asking
Is Waddle-1 on the official RoboDojo leaderboard? Yes. RoboCurve lists Waddle-1 at 81.4% with per-trial artifacts on its November 20, 2026 RoboDojo-RC Tier 1 table.
Does 81% mean the robot finished every task? RoboCurve reports mean rubric progress, not binary success. An 81% average can still include failed stages inside a task.
Can I reproduce the score locally? Only if Waddle Labs publishes the harness, model IDs, and attempt caps. Until then, treat marketing percentages as claims.
How does GPT-6.2 Astra at 65.7% relate to the 65 in Waddle's tweet? Rounding and timing differ. Use RoboCurve's 65.7 figure for the verified frontier row.
Is this the same Waddle as Meta's pet app? No. Meta shipped a consumer journal app called Waddle World in the Philippines. Waddle Labs is a separate robotics agent vendor name collision.
Should I pick Waddle-1 over uni-robot-agent? Different stacks: uni-robot-agent is an open MIT harness you run on simulators today; Waddle-1 is a hosted agent API claiming real-robot RC scores. Choose based on deployment path, not a single percentage.
Historical RoboDojo scores for context
When GPT-6 Astra launched, RoboCurve and simulation leaderboards showed how fast frontier LLM agents could climb embodied tasks without robot-specific pretraining. The jump from single-digit simulation success to forty-plus percent on RC Tier 1 within months is why venture and lab teams keep funding tool-calling arms instead of only vision-language-action monoliths.
The 81.4% row is a step change above the 65.7% capped LLM frontier entry on the same Tier 1 table — but ranking still depends on whether you accept RoboCurve's no-cap regime for vendor full stacks. Outliers on embodied benches often trace to hidden priors (camera placement, object sets, operator hints). RoboCurve's video requirement is the antidote; watch the trials before you repeat the number internally or in a fundraising deck.
What developers should do now
- Open robocurve.org and read Waddle-1's per-trial videos and cost row.
- Read trial videos when they ship — RoboDojo-RC grades staged rubric progress, not binary success alone.
- Request API docs from Waddle Labs if you need zero-shot bimanual control; treat preview terms like any cloud robotics API (latency, safety stops, data retention).
- Compare harness docs side by side with Perceptron Isaac checkpoints if you run YAM arms locally.
- Do not retire simulation benchmarks — Isaac-ready stacks still matter for iteration speed.
Related reading
- uni-robot-agent and the harness catch
- AA-Robotics frontier vs humanoid scores
- Perceptron Isaac 0.5 YAM checkpoints
- Gemini Robotics 2.5 factory demos
- NVIDIA GR00T and agent substrate context
- What is MCP for tool use?
Leaderboard figures follow RoboCurve as of November 20, 2026; social posts may round the same published row.
