Tried @dreamlabs_ai’s API on our own YAM Box from @i2rt_robotics “Pick up the blue cylinder and place it next to the red bottle.” Still a long way from reliable deployment, but this makes us excited about the potential for zero-shot generalization across robot platforms. Here’s how it went 👇

An agent generates trajectories for carrying a kettle and pouring water. Dyna2-style RL handles low-level control with precision and robustness to disturbances—beyond motion planning alone, as Qia Yuan noted. Agentic planning + conditioned RL is a huge frontier.

One thing we learned from simulation: the goal is to transfer skills. We trained manipulation experts in sim, distilled them into vision policies with visual randomization, and deployed them to real robots. The resulting policy handles randomized object placements smoothly across the space, with no need for fixed positions. Demo below ⬇️

One takeaway from #IROS2026# : the core of embodied RSI (recursive self-improvement), is what a robot can already do shapes what it can learn next. Think of how data collection changes as a policy gets better. Once a trained policy can handle most of a task on its own, human operators move past full demonstrations and step in only where it fails, correcting those moments. The data shifts from broad imitation to targeted, high-value snippets that would be far more expensive to source from scratch. Simulation can provide the initial skills that get this process moving. The important design choice is deciding which physical details matter for the skill being learned. A simplified environment can be useful when it preserves the relevant constraints and interactions. Real-world testing then reveals where those assumptions stop holding. That suggests a practical development cycle: Train an initial capability in simulation. • Use it to support more informative data collection. • Study failures to find missing experience or wrong assumptions. • Update the setup and test again. At Axis Robotics, this connects closely to how we think about scaling robot data. Alongside dataset size, we should ask what new situations the data covers and which weaknesses it helps resolve. A useful simulation pipeline should make the next experiment more informative. A useful dataset should make the next policy more capable. The opportunity is to build those two processes together, so each improvement helps create the conditions for the next one.

After injecting user data as a prior, we can train a new expert in ~20 minutes. No complicated reward engineering. No expensive reward-model fine-tuning. Just sparse, simple rewards are enough to adapt the policy across tasks. A scalable way to produce large numbers of high-quality robot experts with very low compute and human cost.

A handful of Axis user demonstrations is enough to bootstrap 95%+ SR RL skills for cross-object grasping and goal-pose tracking across the randomized workspace, training in a single GPU for 3~12H The same general tracking skill can then be quickly fine-tuned for downstream physical requirements, e.g. contact, speed, force.