Robots learn to improve themselves: recursive self-improvement and cheaper action models for physical AI

The ladder nobody has climbed

The most interesting robotics paper this week is not a model release — it is an audit. The authors of Toward Recursive Self-Improvement in Physical AI define recursive self-improvement as a system that improves not only its behavior but the process by which it improves. They lay out a ladder: behavioral improvement, then benefit to later learning, then improvement of the learning procedure itself, and only then true recursion. Their review finds existing physical-AI studies, concentrated in manipulation, support the first three rungs. For the fourth, their verdict is blunt: the physical studies reviewed have yet to isolate gains attributable to recursion. m:…0527

The paper earns that verdict by stating the conditions plainly. Gains carry forward, the authors argue, only when feedback addresses remaining learning needs, the receiving component can actually use the change, its value survives changes in tasks or hardware, and continued practice stays safe and affordable. That is a checklist for any founder evaluating a self-improving robot pitch. And it explains why the physical world is harder than the digital one: physical learning means partial observations, costly trials, changing conditions, and potentially irreversible consequences — against digital agents that can rewrite their own code and check each version with executable tests. The authors' open questions are equally concrete: fair tests a system cannot overfit to, gains that keep stacking up, and safe practice while the procedure itself changes. m:…0527

Cheaper ways to act

If the recursion rung is unclimbed, this week's engineering is about making the lower rungs cheaper. MambaAct attacks the action expert inside flow-matching vision-language-action policies, replacing the transformer-based expert with a compact Mamba-based one plus a lightweight Context Compressor that maps VLM prefix tokens into a compact conditioning signal — no more repeated token-level cross-attention. The authors report an action expert 10 times smaller than the transformer-based expert in pi0.5, flow-matching denoising latency down 64.8% on LIBERO and 58.8% in real-world deployment, while mean LIBERO success rose 1.08 percentage points across three independent fine-tuning seeds. Smaller, faster, and slightly better — the rare combination. m:…8669

SmolVLA-MoE takes the sparse route instead: it integrates a sparse mixture-of-experts into the VLA model's feed-forward layers, so different experts activate for different tasks during robotic action generation. Sparse activation means roughly 500M activated parameters out of about 800M stored, and the authors report 84.25% average success on the LIBERO benchmark — ahead of dense SmolVLA's 82.75% and OpenVLA-7B's 76.5%. Both papers are preprints carrying their authors' own measurements, but they tell the same economic story: the action layer of robot policies is getting dramatically cheaper to run without surrendering success rates. m:…4399

Worlds grounded in skeletons

The other half of the research layer is the world model — the robot's imagination of what happens if it acts. SkeLAM decomposes latent actions into a skeleton-guided kinematic component and a residual component carrying complementary visual dynamics, trained with cross-modal swap reconstruction so that RGB-derived and skeleton-derived representations become interchangeable during video reconstruction. On the DROID dataset, end-effector ADE fell 19.5% and LPIPS 11.1% versus DreamDojo, qualitative results show zero-shot human-to-robot motion transfer in generated rollouts, and — the detail practitioners will care about — policy evaluation showed predicted success rates agreeing with measurements from real robot executions. m:…3374

The unclimbed rung

Read together, the week looks like infrastructure for a bet that is still unproven. Cheaper action models and better-grounded world models make robot learning faster and cheaper — exactly what you would need to start climbing the self-improvement ladder. But the audit stands: the recursion rung is unclimbed, and the ladder's own authors say no physical study has isolated recursion gains. For founders, the honest read is to invest in the measurement and world-model layers and to ask hard questions of anyone selling the top rung before it exists. This suggests the physical-AI companies to watch are the ones building better simulators and cheaper policies, not the ones promising robots that redesign themselves. m:…0527 m:…3374 m:…8669