Recursia Beats Nvidia ASPIRE by 10 Points on Long-Horizon Robot Tasks

Recursia, a method within the Embodied RSI paradigm, outperforms Nvidia ASPIRE by 10 points on long, multi-step robot tasks. The Chinese outlet Jiqizhixin reported the result. The developers are aiming for zero-shot generalization: a manipulator that can carry out a sequence of actions in scenes it did not see during training. This is where today's robots break down most often, even when their individual skills are well polished.

Task T3: why clearing a table is hard

To illustrate, the authors point to the long task T3, performed on a real robotic arm. The robot places a bowl on a plate, then picks up a cup and puts it into the bowl that was just moved. For a human, this takes a couple of seconds. For a robot, the difficulty is that the second step happens in a scene the first step has changed. The bowl is no longer where it started, and the positions of the objects differ from what the model saw at the outset.

Such errors accumulate. Recognition, grasping, and placement may each work with high reliability on their own, but over a chain of several steps small inaccuracies multiply, and the policy ends up in states that were not in the training data. That is why long tasks have become the main test of generalization: short demos hide a lot, while long ones show where a system falls apart.

What is known about Recursia and RSI

In robotics, RSI is usually expanded as recursive self-improvement: a system uses its own experience and results to get better. Embodied RSI carries this idea over to physical agents. The authors themselves are sober about it: the article states outright that the "RSI moment" for robots is still very far off. They present Recursia as a step in that direction, not as its arrival.

The available excerpt of the publication yields the following:

• method: Recursia, Embodied RSI paradigm;
• goal: zero-shot generalization for robots;
• comparison: Nvidia ASPIRE;
• result: a 10-point lead on long tasks;
• validation: a real robot, with task T3 as the example;
• source: Jiqizhixin, October 14, 2026.

The article leaves important questions open. It is unclear what the "points" measure (probably the share of successful runs, but this is not confirmed). The sample size, the number of long tasks, the length of the chains, and exactly how ASPIRE was run as a baseline are all unknown. There is no data on how much simulation and how many real-world attempts the self-improvement required.

How to read results like this

For engineers and investors who follow robotic manipulators, the main practical takeaway is this: focus on long tasks in which the scene state changes, not on demos of individual skills. That is where real generalization shows. Before drawing conclusions about Recursia's superiority, wait for a paper or preprint and check three things: the metric, the number of runs on real hardware, and whether the comparison with ASPIRE is fair, using the same tasks and the same conditions.

Ten points on a benchmark with a few dozen runs and ten points over hundreds of attempts across different scenes are different claims. Until the details are published, this is a signal about direction, not a proven breakthrough.