Skill-Space Shooting for Autonomous Robot Policy Improvement

What it claims

Skill-space shooting is a DAgger-style improvement loop in which a library of separately taught short-horizon skills (primitives) stands in for the human expert. When a learned detector predicts that the task policy is about to fail, a VLM judge picks a skill and runs it from the state the robot has actually reached. Only verified repairs become training data. The authors report repeated gains in unaided policy performance across four real-world tasks (Stack-3, Drawer, Sweeping, Coffee), including one policy that started with no observed task successes. They also report that shared skill data reduces the human effort needed to adapt skills to new primitives and to improve on a new task.

How it works

Inputs are a behavior-cloned task policy and a skill library. The library is seeded by demonstrating each task stage independently, and skills can come from earlier tasks.

• When to intervene. A value model Vϕ scores the current observation together with the policy's proposed 16-step action chunk. It uses DINOv2 image features and encoded action tokens to condition a flow-matching head, which predicts the post-chunk robot state and a scalar value. Training targets are +0.5 for successful execution and −0.5 starting one action horizon before a recorded failure, so the detector fires early enough for a skill to recover. The threshold is τ=0 on every task, with no sweep. Each task's detector is trained on initial labeled episodes and then frozen.
• What to try. Gemini 2.5 Pro receives the scene, the task instruction and the skill descriptions through a fixed prompt shared across tasks. It selects a skill, for example spotting that the sponge was never grasped before wiping.
• Verification. The value model monitors the skill while it runs. After c consecutive nonnegative checks (c=8, or c=16 for Coffee) or M=100 actions, the VLM judges whether the stage was repaired. A failed check retries the same skill. A rollout is discarded after three failed retries.
• Learning. The training data keeps only accepted, skill-executed segments from rollouts that the judge confirms as complete. Each segment runs from invocation to handback and is relabeled with the task instruction. The policy's own actions and any rejected trials are dropped. Retained repairs Rk are aggregated with the initial demos D0 for each iteration. Training can start before the policy's first end-to-end success.

flowchart TD
  P[Task policy proposes chunk] --> V{Value < 0?}
  V -- no --> P
  V -- yes --> J[VLM selects skill]
  J --> S[Skill executes]
  S --> C{VLM: stage repaired?}
  C -- no, retries left --> S
  C -- budget exhausted --> X[Discard rollout]
  C -- yes --> P
  P --> E{Task complete?}
  E -- yes --> D[Keep repair segments + D0 → retrain]
  D --> P

Evidence

The fetched text cuts off in the middle of the method section, before any results table. The only result claims available come from the abstract and introduction: improvement over successive iterations on four real-world tasks, including from a policy with zero initial successes, an intervention analysis showing that the library supplies useful repairs at failure states, and reduced teaching effort when skill data is shared. No success rates, iteration counts or baselines are visible.

Limits

• Skill selection is mostly trivial in the experiments. Each task stage has one designated primitive, so identifying the failing stage determines the skill. The VLM's skill-selection ability is therefore largely untested.
• Humans are still in the loop. The loop still needs human scene resets. It also needs per-task labeled episodes to train the failure detector and stage-wise demonstrations to seed the skills.
• Reliance on a proprietary model. The design depends on a closed model (Gemini 2.5 Pro) both for skill selection and for confirming task completion.
• Truncated and unreviewed. Quantitative results, ablations and the authors' stated limitations are not in the fetched text, so they need to be checked in the full paper before relying on the claims. This is an arXiv v1 preprint dated 29 Sep 2026. Videos are at skill-space-shooting.github.io. The page mentions no code release.