Robots learn new skills without fresh demos — InSight steers actions piece by piece
By splitting demonstrations into small “primitives” and looping successful attempts back into training, InSight composes long tasks from learned moves. Paired with AGORA’s archive-grounded test and Composio’s 1,000+ tool developer toolkit, agents that act and reason get a practical boost.
One-Line Summary
Agents move beyond chat: robots self-learn skills from atomic moves, while a new benchmark tests document-grounded reasoning and open tooling wires agents to real apps.
Research Papers
InSight: self-guided manipulation via steerable action primitives
InSight is a framework that lets robot learning systems pick up new manipulation skills by steering low-level “primitive” actions like “move to bowl,” “lift upward,” or “pour.” Unlike typical Vision-Language-Action (VLA) models that stay limited to what they saw in demonstrations, InSight practices missing moves and adds successful attempts back into training, enabling new tasks without fresh human demonstrations. 1
It works in two stages. First, an automated segmentation pipeline splits demonstrations into labeled primitives using a Vision-Language Model (VLM) to decompose plans and end-effector poses to tag motions. Second, a VLM-guided data flywheel spots missing primitives, proposes low-level controls, attempts them autonomously, and auto-labels and stores successful trials for future VLA training. 1
The authors evaluate InSight in simulation and on real robots across tasks including block flipping, drawer closing, sweeping, twisting, and pouring—without any human demonstrations of those target skills. Once learned, these primitives can be composed to execute novel, long-horizon tasks without additional human demos. 2
By making VLA policies steerable at the primitive level, InSight points to a practical path for continual skill acquisition: build reusable action pieces and keep expanding the set as tasks demand. What to watch: robustness of VLM guidance in varied scenes and safe handling of failed self-practice attempts. 1
AGORA sets a hard test for workplace document agents
AGORA is a benchmark for agents that must find sparse evidence across large, messy workplace archives and compute an answer, instead of replying from memory. It pairs 362 questions with eight domain collections spanning 9,664 authentic documents and 372 million tokens—far beyond any model’s context window—so agents must search and plan deliberately rather than read everything. 3
Built with an agentic pipeline that combines cross-document task synthesis, leakage-preventing obfuscation, and difficulty filtering, AGORA shows the task is far from solved: across eight models, the strongest achieves only 59.4% accuracy with notable domain variation. For teams building Retrieval-Augmented Generation (RAG) assistants and workplace agents, these numbers set a concrete baseline and highlight the need for deliberate exploration, tool use, and claim-by-claim checking. 3
Open Source & Repos
Composio aggregates 1,000+ tools for building AI agents
Composio is a developer Software Development Kit (SDK) and toolkit that helps AI agents turn intent into action by connecting to 1,000+ toolkits, providing tool search, context management, authentication, and a sandboxed workbench. It targets teams wiring agents into calendars, CRMs, and internal apps where controlled execution matters. 4
A recent prerelease of its Command-Line Interface (CLI), @composio/cli 0.2.32-beta.273 on 2026-06-24, adopts the Node 24 and pnpm 11 toolchain. As agents gain capabilities, stable tool interfaces and sandboxing become critical; Composio positions itself as the glue layer for that integration. 4
Why It Matters
InSight shows how to grow a robot’s skill set from atomic moves without new human teaching, AGORA quantifies how far document-grounded reasoning still has to go (best accuracy 59.4%), and Composio builds the plumbing that connects agents to real tools—together, they push agentic AI from talk toward reliable action. 3
This Week to Try
- InSight demo and paper: skim the arXiv page and watch clips of primitive compositions in action: https://arxiv.org/abs/2606.24884
- AGORA benchmark: read the paper’s task design and sample questions to stress-test your agent mental model: https://arxiv.org/abs/2606.24526
- Composio quickstart: browse the GitHub README and try the CLI prerelease in a test project: https://github.com/ComposioHQ/composio
Comments (0)