Skip to content
RobothecaBeta
Log in

Flex-π

  • AI model
  • Humanoids and robot learning

A 6B-parameter world-action model for robot manipulation that jointly predicts future video, 3D pointmaps, semantic features and actions, runnable at different compute levels from one checkpoint.

Source on GitHub: github.com/geyan21/flex-pi (opens in a new tab). Facts read at commit 20c1b2b (opens in a new tab).

Licenses

Code
MIT (opens in a new tab)

Read by Robotheca from LICENSE (opens in a new tab) (MIT) at commit 20c1b2b (opens in a new tab), on 16 Oct 2026.

Commercial use
Yes

What it does

  • Predicts future RGB frames, 3D pointmaps, DINOv3 semantic features and robot actions during training, using a Mixture-of-Transformers architecture.
  • Trained with random stream dropout so a single checkpoint can run any subset of input and generated streams.
  • Action-only inference is the fast path; full joint generation of visual futures is slower but more accurate.
  • Evaluated on RoboTwin 2.0 and LIBERO simulation benchmarks, and deployed on a bimanual YAM real robot via a WebSocket policy server.
  • Checkpoints and datasets are published on Hugging Face; training and evaluation launchers are provided per benchmark.
  • Inference can be sped up training-free with torch.compile, with optional TensorRT engines for the joint path.

Written by Robotheca from the project's README (opens in a new tab) at commit 20c1b2b.

Getting started

Difficulty
Advanced
Hardware it needs, as the project states it
Linux x86_64 with CUDA 12.8 and Python 3.10. Inference and deployment need 16–26 GB memory on one RTX 4090 or 5090 (peaks measured on an RTX 5090). Training needs 80 GB per GPU on 4–8 A100 80GB, H100 or H200 GPUs; the launchers assume 8 GPUs and 4 is the stated floor.

Maintenance

Status
Active: a commit in the last 90 days
Newest commit
23 Aug 2026 on main
Latest release
None published on GitHub

Checked on GitHub on 16 Oct 2026.

Links

Robots that can run it

No robot on Robotheca is linked to it yet. Links come from a robot's robot.yaml, from its dependency files read as text, or from Robotheca's editors.

Robotheca lists AI models as a reference: nothing is hosted or run here, and every fact links to where it was read.