Research · October 2026
R2SI is a two-layer robot agent. The inner Robot Agent acts. The outer Recursion Agent improves it — revising Skills and training the reactive model, round after round.
39.35%SR46.58Score
#1 Overall35.83%SR42.95Score
Generalization15.50%SR23.51Score
Precision34.25%SR51.48Score
Long-Horizon63.67%SR65.78Score
Memory47.50%SR49.16Score
OpenThe architecture, push_T and packing, end to end.
R2SI is built from two agents. The inner layer executes. The outer layer improves. Every update must pass its tests before it ships.
Inner · per decision
Decides from multi-view RGB images, robot proprioception, the task description and episode memory, and acts through the Skills it loads.
Calling the reactive model is itself a Skill: the model executes one action chunk, then hands control back, and the Robot Agent decides the next step from fresh observations.
Outer · per round
Reviews execution records across rounds, revises Skills, and applies them to the inner layer as Harness updates.
For tasks that use the reactive model, it also turns execution records into training data and trains the model. Updates that pass their tests are released automatically and versioned.
Official native step limits, RGB and proprioception only. R2SI runs on GPT-6-Astra — the same VLM as the reference agent on the leaderboard — so the gap is what the harness adds.
Overall score 0–100
Green gain: score points over GPT-6-Astra (same VLM).
In-budget success rate %
Green gain: percentage points over GPT-6-Astra (same VLM).
Generalization
42.95+6.57 over VPP2. Standard and random layouts average 35.83% success, under the official native step limits.
Memory
65.7863.67% success on tasks that depend on earlier observations. +22.74 score over GPT-6-Astra in the reference harness.
Same VLM
30 / 42Tasks where R2SI beats GPT-6-Astra in the reference harness. Memory +22.74, Long-Horizon +30.03, Precision +10.86.
All 42 tasks. R2SI: official leaderboard, 2026.10.17, with all 54 variants aggregated by task. Baselines: official leaderboard, 2026.10.14.
Remove one component, keep everything else fixed.
The outer layer improves the inner one along two paths. Teal is the Skill path (ΔHarness); violet is the reactive-model path (Δθ).
The inner layer acts through Skills; the outer layer keeps improving the Skills and the reactive model.
A full stack_blocks success in 28 decisions. Three camera views in, one typed tool call out — and when the planner rejects a pose, the agent picks another instead of stalling.
Push a T-shaped block onto the gray T pad within 600 native steps. The outer loop rewrites the Skill until the run fits the budget with room to spare — then the frozen Skill solves layouts it has never seen.
Push a ruler against three blocks until they line up, within 200 native steps. Every candidate Skill is tested on the same 8 layouts. The gate rejects three of them — one slower than its predecessor — and promotes v4, the first to finish all eight in budget.
01 · Memory → Skill
active-approach/SKILL.md
02 · Tool · test
gate: 8 / 8 pass → release
03 · Release
The passing version replaces the old one in the inner layer automatically and is recorded as a new Harness version. Mean steps fall from 215 at v1 to 140 at v4.
v0 → v4
# Push the ruler with one open hand; fewer repeated moves The ruler rests on the table. Push it toward the three blocks with the pads or outside of an open gripper. Do not grasp, close on, or lift the ruler. Goal: align the contact edges and orientation of all three blocks, return the robot to its start pose, within ~200 native steps. 1. Decide the contact plan once. From head and wrist RGB, find the ruler, the blocks and the reachable area behind the ruler; save both arms' start poses as the home target. 2. Set the pose first, then descend. Combine approach position and target orientation in one end-effector move; if the planner rejects it, adjust one necessary intermediate pose. 3. Descend behind the ruler. If the first short push does not move it, stop and check the finger height in wrist RGB before trying again; do not keep lengthening the push. 4. After one confirmed contact, push in effective centimetre-scale strokes, not repeated 5 mm probes. 5. Check each block's edge against the ruler; correct only the block that is still tilted. 6. Clear the fingers before turning the wrist, then return home in a single move.
When the task switches to packing, the Recursion Agent loads the Skill that calls the reactive model. The Robot Agent calls the model from its observations; the model executes one action chunk and hands control back; the Robot Agent decides the next step from new observations — and so on.
The Recursion Agent organises the execution records into training data and trains the reactive model. The updated model goes back to the inner layer to be called again — so it both executes and keeps improving under the outer agent’s training.
R2SI runs the whole RoboDojo suite — 42 bimanual tasks and 54 variants across five ability groups — with one Robot Agent and one toolset.
The inner layer acts through Skills — and calling the reactive model is one of them. The outer layer keeps improving both the Skills and the reactive model. Trial and error settles into reusable capability.
This is R2SI.