TinyAstral: Transferring Self-Evolving Harnesses to Tiny Generalist Manipulation Agents

TinyAstral: Transferring Self-Evolving Harnesses to Tiny Generalist Manipulation Agents

Robot manipulation systems are often described as if the model alone decides whether a task succeeds. In practice, a vision-language model operates through a harness: the surrounding software that exposes tools, interprets observations, stores task information, checks outcomes, and handles failures. The authors of TinyAstral argue that improving this layer can be as important as improving the planner itself—and that those improvements can be transferred to a much smaller model.

The problem: capable teachers are expensive

Large models can reason about which object to manipulate, where to grasp it, and when a task is complete. They can also interpret tool feedback and recover after a missed grasp or misplaced object. However, running a large teacher for every interaction consumes substantial inference compute. A compact model is cheaper, but may repeat failed actions, misunderstand feedback, or terminate prematurely.

TinyAstral addresses this tension with a two-stage process. First, the system improves a large teacher’s manipulation harness while keeping the teacher’s weights frozen. Second, it trains a smaller student from trajectories produced by that improved system. The student then replaces the teacher inside the same frozen harness.

Stage one: evolve the harness, not the model

The authors use a master system of frontier language-model agents to inspect logged trajectories, classify failures, and propose changes to tools, feedback, memory, and recovery logic. Candidate edits are tested against the current harness on the same task-and-seed pairs. An edit is accepted only when it passes pre-registered paired statistical tests designed to show improvement on the targeted cases without unacceptable regressions elsewhere.

This constraint is important. The teacher’s model weights do not change during harness evolution, and the master does not directly act in the environment. The paper also describes safeguards intended to prevent task-specific overfitting, including disjoint selection and evaluation seeds, explicit triggers for new components, and audits checking that untriggered behavior remains unchanged.

Across the measured harness lineage, the frozen teacher’s strict success on 500 RoboTwin episodes increased from 33.4% at an early version to 63.6% in the final version. The gains were not limited to that teacher: DeepSeek-V4.1-Flash improved from 32.8% to 63.4%, while an untrained 4B planner improved from 14.6% to 41.2% under the same sequence of harness versions.

Stage two: distill reasoning and recovery

After evolution ends, the final harness is frozen. The teacher generates trajectories containing its reasoning, tool calls, feedback, and recovery behavior. The student is trained to predict both the rationale and the next tool call, rather than only imitating the final action.

The training set also includes deliberately perturbed recovery branches. The researchers replay successful episodes, inject failures such as slipping objects or moved targets, and retain branches where the teacher recovers and still achieves strict success. Offline real-robot data is added to expose the student to real images and tool-use patterns.

The final 4B student is based on Qwen3.5-4B and runs in the evolved harness. On RoboTwin it reaches 62.8%, close to the 63.6% teacher. It also stays close on Guava, at 93.6% versus 94.7%, and on LIBERO-PRO, at 78.3% versus 80.0%. The authors report that students trained from four successive harness snapshots improve alongside their teachers, rising from 36.2% to 62.8%. This is the paper’s central evidence that harness improvements can survive distillation.

Results beyond the training domains

TinyAstral-4B was also evaluated on SimplerEnv, VLABench, and EB-Manipulation, which the authors say were not used for RSI or training. The student achieved 66.6% and 88.5% on the two SimplerEnv settings, 41.2% on VLABench, and 27.6% on EB-Manipulation. On these tests, its results were within 5 percentage points of its teacher and within 2.4 points of the reported 552B planner in the same harness.

The smaller models show a mixed picture. TinyAstral-0.8B reaches 61.8% on RoboTwin and 93.8% on Guava, but falls to 54.7% on LIBERO-PRO, compared with 78.3% for the 4B student. The authors therefore show that the approach can produce very small agents, but not that reducing model size is cost-free across every task family.

What mattered in the ablations?

Preserving the teacher’s reasoning had the clearest measured effect. Removing reasoning while leaving the other training samples unchanged reduced success by 19.3 points on LIBERO-PRO, 11.2 points on SimplerEnv Google VM, and 10 points on VLABench. The authors interpret this as evidence that the rationale helps the student connect visual feedback with its next corrective action.

Real-robot supervision substantially improved offline next-call accuracy on real observations, but it did not produce a statistically detectable success improvement across the seven simulated settings tested. That distinction matters: better transfer to real-looking inputs was demonstrated in an offline measure, not necessarily in broad closed-loop task success.

Compute savings and practical implications

On 20 shared SimplerEnv episodes, the deployed 4B student used 25.3 accelerator-seconds per episode, compared with 123.8 for the teacher and 249.6 for the other large planner measured by the authors. In that deployment measurement, the student therefore used about one fifth of the teacher’s accelerator time.

For practitioners, the result suggests a useful development pattern: invest expensive model-based reasoning in improving the interaction layer, then distill the resulting behavior into a cheaper policy. This could be attractive when tool reliability, verification, and recovery are bottlenecks rather than raw perception alone.

Limitations and open questions

The evidence remains bounded by the authors’ environments, harness, data, and evaluation protocols. Several competitor results use different capabilities or stopping rules, so headline comparisons should be read carefully. The final system also relies on a substantial engineered harness, including perception and motion-planning services, meaning the 4B model is not a standalone robot controller.

More broadly, the paper demonstrates transfer across the reported benchmarks and snapshots, not universal transfer to arbitrary robots or tasks. Real-robot evidence is limited to closed-loop examples and offline supervision details in the supplied text. Future work will need to test whether automatically evolved recovery logic remains safe, maintainable, and effective as hardware, tools, and task distributions change.


Sources

AI-assisted research summary. Check the original paper for complete methods, results, and limitations.

Avatar photo

Tech News

Jeff is a passionate blog writer who shares clear, practical insights on technology, digital trends and AI industries. With a focus on simplicity and real-world experience, his writing helps readers understand complex topics in an accessible way. Through his blog, Jeff aims to inform, educate, and inspire curiosity, always valuing clarity, reliability, and continuous learning.

↑ ↓ navigate Enter open Esc close