Is DreamFAR state of the art?Pending benchmarks.Announced Oct 29, 2026; no independent results yet.

The details so far

Lab

Amazon

Announced

Oct 29, 2026

API

Not yet

Open weights

Announced, not yet published · Code and model weights announced as coming soon; none published yet

Parameters

1.9B dynamics transformer plus a 46M encoder and 50M decoder tokenizer (paper figures; weights not yet released)unsure

Type

Action-conditioned robot world model (latent flow matching) for bimanual manipulation

Architecture

Causal transformer with 28 blocks of hidden size 2048 that predicts the next latent frame from the action, recent frames and contact-based memory keyframes retrieved by end-effector pose

Tokenizer

Unit-sphere tokenizer fine-tuned from a pretrained SSDD-M checkpoint; each 256 x 256 camera view becomes a 16 x 16 grid of 16-dimensional latents

Training data

Internal bimanual teleoperation dataset on YAM robots (two 6-DoF arms with parallel grippers, three camera views, 14-D actions at 10 Hz); dynamics trained from scratch, no internet-video pretraining

Training compute

256 H200 GPUs for six days

Context

5-second window of 50 frames plus eight memory slots and one learned sink frame

Control inputs

Recorded actions, a policy, or a person via mouse, keyboard, webcam hand tracking or GELLO leader arms

Live demo

Public web demo on the project page; one driver at a time in 4-minute sessions

Paper

14-page PDF on the project page; no arXiv listing

Authors

Amazon FAR, with co-authors from Cornell University, University of Maryland and University of Pennsylvania

Release

Project page and X announcement Oct 29, 2026

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: real-time interaction at 10 Hz (10+ FPS on a commercial GPU), 5+ minutes of autoregressive rollout, and generalization across 1,000+ tasks.

Lab claim: with every model trained on the public ABC dataset, DreamFAR averages 16.4 dB PSNR over 60-second rollouts on 1,000 held-out episodes, versus 15.8 for Cosmos-Predict 2.5, 14.5 for WEAVER and 11.0 for IWS, with the lowest LPIPS in every 10-second bin.

Lab claim: policy task progress inside DreamFAR correlates with the real robot at Pearson r = 0.71 across two policy checkpoints and three tasks; it is optimistic in absolute terms.

Lab claim: fine-tuning pi0 on 75 real plus 75 DreamFAR-teleoperated demonstrations of placing chip bags in a basket nearly doubles real-only task progress, to 63%.

Sources