Robots are learning to watch, look around, and soon grasp — but the honest read says not yet
Two papers tighten the video-to-action loop, a third asks whether GPT-6-Astra is truly embodied, and South Korea funds a 25-degree-of-freedom hand. The loop from model to metal is closing.
- In a preprint, BridgeWAM decouples video and action experts through 32 latent bridge queries, reporting 98.25% success on LIBERO, 28.72% fewer single-sample inference FLOPs than Fast-WAM, and 77.1 ms latency on a real robot. Self-reported m:…1616
- In a preprint, ACT3D trains vision-language models with reinforcement learning to actively explore 3D from single first-person views, reaching 53.2% average accuracy across six 3D benchmarks — a 22.0% gain over the base Qwen model — while using 50–70% fewer input frames than open-source 3D methods. Self-reported m:…2652
- In a preprint, the paper's authors assess that GPT-6-Astra has crossed the threshold of embodied competence but not practical embodied intelligence: demonstrations for in-context learning are expensive to collect and the System-2/System-1 coordination is costly. Self-reported m:…0771
- Tesollo says it was selected to lead a project under South Korea's 2026 Robot Industry Technology Development Program to build a 25-degree-of-freedom robotic hand with multimodal sensing and on-device AI: about KRW 5.2 billion total, roughly KRW 4 billion in government funding, August 2026 through December 2029. Self-reported m:…3349
Video and action, decoupled
World-action models usually wire video and action experts together through dense layer-wise attention, which exposes action prediction to irrelevant visual features. A preprint posted this week, BridgeWAM, breaks that coupling: a Bridge-of-Experts aggregates visual, language, and proprioceptive priors through 32 latent bridge queries, so action gradients guide the visual representation through a shared bottleneck instead of a dense mesh. m:…1616
The reported numbers: 98.25% success on LIBERO, 72.24% pooled on LIBERO-Plus, 28.72% fewer single-sample inference FLOPs than Fast-WAM — a 2.47-fold speedup — and 77.1 ms latency on a real robot. The model is 5.06B parameters, with a two-layer action expert sitting on a 30-layer expert. m:…1616
The latency figure is the one to watch for anyone funding robot stacks. Sub-100 ms real-robot loop times put video-to-action models in the control regime, not the planning regime. m:…1616
Learning to look around
A second preprint, ACT3D, attacks the other half of embodiment: perception that moves. It trains vision-language models with reinforcement learning to actively explore 3D environments from single first-person views — preserving the models' native visual encoding rather than pre-rendering 3D — with rewards driven mainly by final-answer correctness. m:…2652
Result: 53.2% average accuracy across six 3D benchmarks, a 22.0% improvement over the base Qwen model, surpassing Gemini 3.6 Flash — while using 50 to 70% fewer input frames than existing open-source 3D methods. The gains scale with how much exploration a question demands: from 0.6% on initial-view questions to 16.3% where extensive exploration is required. m:…2652
For investors, this is the perception story underneath manipulation. A robot that searches for occluded targets, avoids unnecessary exploration, and acquires complementary viewpoints without being told which viewpoints to visit is doing something closer to embodied reasoning than pattern completion. m:…2652
The honest counterweight
Then comes the paper that asks the uncomfortable question directly: has GPT-6-Astra achieved embodied intelligence? The model was evaluated across 12 representative benchmarks — tabletop manipulation, humanoids, mobile platforms, dexterous hands — and on 8 real robot configurations, and shows high performance via in-context learning coordinated between a high-level System 2 reasoner and a low-level System 1 controller. m:…0771
The paper's authors assess that GPT-6-Astra has crossed the threshold of embodied competence — but not the threshold of practical embodied intelligence. The demonstrations required for in-context learning are expensive to collect, the System-2/System-1 coordination is costly, and execution time is constrained: the authors note it is implausible to trade longer execution for higher success rates under physical constraints. m:…0771
That is the framing worth keeping: benchmarks that permit deliberation without a running world clock flatter embodied models. The prototype for next-generation benchmarks in the paper derives tasks from real-world needs, provides no in-domain demonstrations, and judges agents on a world clock that runs during deliberation. m:…0771
From model to metal: a funded hand
While the papers argue about thresholds, hardware funding is moving. Tesollo says it has been selected to lead an R&D project under South Korea's 2026 Robot Industry Technology Development Program, run by the Ministry of Trade, Industry and Resources — a 25-degree-of-freedom robotic hand with multimodal sensing and on-device AI, built alongside partners Device Corporation, Hanyang University, and the Korea Institute of Robotics and Technology Convergence. m:…3349
The budget, per the company: about KRW 5.2 billion total, including roughly KRW 4 billion in government funding, running August 2026 through December 2029, targeting technology readiness level 7 or higher. Tesollo plans commercial solutions for humanoid manufacturers and industrial robotics companies, and says the on-device AI platform is designed for low power consumption and high efficiency. m:…3349
Twenty-five degrees of freedom is roughly human-hand territory, and the pairing is telling: on-device AI sits at the center of the specification, not the cloud. If dexterity is where embodiment gets scored, a funded 25-DOF program with a 2029 horizon and commercial intent is the concrete side of the loop the papers describe. m:…3349
What is still unresolved
- All three papers are preprints; benchmark numbers come from the authors' own evaluations.
- GPT-6-Astra's competence-versus-intelligence thresholds are the authors' proposed definitions, not an industry standard.
- Tesollo's timeline and budget are company-reported; a TRL-7 hand by 2029 is a plan, not a delivered device.
- Whether BridgeWAM's 77.1 ms latency and ACT3D's frame efficiency hold outside the authors' test rigs.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.