TBNAI
Long-WAM: Scaling the Context of World-Action Models
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 06/10/2026, 20:00:00 UTC
Briefing created: 04/11/2026, 11:22:07 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 08/10/2026, 03:18:25 UTC.
Briefings
Long-WAM is a model-system framework designed to scale causal world-action models under real-time control by leveraging long video histories. The study found that video foundation pretraining autoregressively (AR) significantly outperforms bidirectional pretraining, boosting success rates from 63.3% to 78.7% on RoboCasa GR-1 when context increased from 0.0 to 19.2 seconds. The framework achieves best results on LIBERO-Long, RoboTwin 2.0, and DOMINO, while enabling deployment on consumer-grade hardware like the RTX 5090 with 107.4 ms latency per action chunk.
Key points
- Long-WAM uses video foundation pretraining autoregressively (AR) to scale causal world-action models, achieving 78.7% success on RoboCasa GR-1 with 19.2 seconds of context compared to 63.3% with bidirectional pretraining.
- The framework supports streaming observation encoding and asynchronous execution, allowing deployment on consumer hardware like the RTX 5090 with 107.4 ms latency per action chunk.
Context
This research addresses real-time robot control challenges where processing video history can delay actions. Long-WAM provides a framework for scaling causal world-action models under these constraints, specifically comparing AR pretraining against bidirectional pretraining to optimize success rates in dynamic environments.
Limitations
Results are limited to specific hardware (RTX 5090, DGX Spark, Jetson AGX Thor) and robot models (RoboCasa GR-1, Unitree G1, YAM). Success rates vary by task (e.g., 95% on cup stacking vs. none for Pi0.5), and the study does not provide causal claims or generalizable forecasts beyond the tested conditions.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 04/11/2026, 11:22:07 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeiavipm6enc6pye6dslgu22bnlztowd63yvbr46hmiindl2so7cor4