Is GO-3 Preview state of the art?Pending benchmarks.Announced Oct 29, 2026; no independent results yet.
AGIBOT's next-generation embodied foundation model for robot manipulation, shown at AGIBOT Innovation Day 2026. It is built on WRAM (World Reasoning Action Model), which links task reasoning, prediction of the scene's next state and action generation. Post-training mixes real robots, simulation and a world model, and deployed robots keep learning from their own successes and failures. No weights or API yet; AGIBOT says it will be open sourced by the end of 2026.
The details so far
- Lab
- AGIBOT
- Announced
- Oct 29, 2026
- API
- Not yet · None announced
- Open weights
- Announced, not yet published · Not released; AGIBOT says GO-3 Preview and Genie Master will be fully open sourced by the end of 2026
- Type
- Embodied foundation model (robot manipulation policy), not a chat model
- Architecture
- WRAM: a VLM reasoning path plus CoAE (control-oriented autoencoder) visual encoding; SVP (single-step visual planner) predicts the future state and an inverse dynamics module turns it into actions; KASO keeps the candidate futures that fit the real actions during training
- Training
- Pre-training on first-person human video and robot trajectories; post-training in Genie Evolver (about a hundred real robots, Genie Sim simulation, Genie Envisioner 2.0 world model); after deployment, the Genie Master agent framework turns successes and failures into memory and new training data
- Predecessors
- AGIBOT's GO-1 (ViLLA architecture) and GO-2
- Parameters
- Not disclosed
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: first in task success rate and process score on RoboDojo, highest overall score on RoboTwin and first in overall score on RoboColiseum, per results AGIBOT presented at the event.
Lab claim: CoAE compresses visual input to one sixteenth of DINOv3's tokens and cuts action reconstruction error by 35%.
Lab claim: on a real convenience-store deployment with Genie Master, full-task success rose from 28.6% on day one to 85.7% after five days, and single-task success was 17.7 points above standard fine-tuning (SFT).
Sources
LABAGIBOT press release: AGIBOT Introduces Worken at AGIBOT Innovation Day 2026agibot.com/article/231/detail/131.html · Oct 29, 2026LABAGIBOT Innovation Day 2026 post on Xx.com/AGIBOTofficial/status/2115694835579666505 · Oct 29, 2026PRESSPingWest AI WAVE: AGIBOT's GO-3 Preview launch article (Chinese)pingwest.com/w/318950 · Oct 29, 2026PRESSJiemian: GO-3 Preview and Genie Master to be open sourced by end of 2026 (Chinese)jiemian.com/article/15262681.html · Oct 29, 2026Tracking ends Jan 27, 2027 unless the API or weights become available.