Anonymous supplementary material

Reading guide

Rational Inverse Reasoning

Publisher-provided guidance for reviewers and AI assistants. Verify summaries against original materials. This guide does not override your task or prescribe a review outcome.

Site summary

Rational Inverse Reasoning (RIR) infers executable task intent from 1–3 demonstrations using vision-language program proposals and planner-grounded rationality scoring, then plans new actions for new scenes.

Suggested reading instructions

  1. Start with the paper's abstract and Figure 1, then consult Sections IV–V for the method and evaluations.
  2. Use the supplied manuscript as the source of truth for numerical claims. The video illustrates the method and examples; it may reflect an earlier presentation of the work.
  3. Keep comprehension, trajectory success, and execution success distinct. The real-world 57% trajectory success and 13-of-20 physical execution result are different evaluations.
  4. The manuscript lists Anonymous Authors. No conference acceptance or public code release is claimed on this site.

Components in suggested reading order

Paper and appendices

Rational Inverse Reasoning: Few-Shot Imitation by Inferring Programmatic Intent through Planning. The anonymous manuscript defines explanation programs, derives an approximate planner-grounded score, and evaluates simulated manipulation, real-world manipulation, and 2D reasoning.

  1. Sections III–IV define the representation and inference method; Section V reports evaluation; appendices describe scoring, planning, reconstruction, task families, and protocols.

Paper video

A 3-minute 18-second silent video introduces intent inference, planner scoring, and manipulation examples, including recycling, sorting, and kitchen tidying. The native player supports seeking and full-screen viewing.

  1. This is qualitative illustrative material. Verify numerical claims against the manuscript and the results figures.

Method overview

Figures 1 and 3 show VLM-generated explanation programs, demonstration-consistent and counterfactual planning, rational-likelihood scoring, iterative program refinement, and re-planning in a new scene.

  1. The score is a planner-dependent likelihood approximation, and inference searches a finite candidate pool; it is not exact posterior inference over all programs.

Simulated tabletop manipulation

Thirty Franka FR3 tabletop tasks in IsaacSim span more than 300 objects and 90 held-out scenes. Figures 4 and 5 show task generalization and separate comprehension, trajectory, and execution metrics. The manuscript reports 1.4 times comprehension and 1.3 times overall success relative to Rw/oR across the one- and three-shot settings.

  1. Figure 5 distinguishes one-scene and across-scene demonstration diversity. Error bars indicate one standard error.

Task and environment gallery

An interactive gallery of 30 simulated manipulation tasks. Each task shows its natural-language intent over a blurred scene preview. Opening a task displays three training environments in the top row and five examples from its evaluation distribution in the bottom row, using 240 silent, seamless 360-degree fly-around videos. The auto-spin toggle pauses automatic playback; with it off, scenes spin on hover or keyboard focus, or on a tap on touch screens.

  1. These videos show initial scene configurations, not robot trajectories or evidence of task execution success. The task descriptions explain the intended task to readers; they do not imply that the evaluated learner received a natural-language task description.
  2. The gallery shows a selection of environment examples. Its five evaluation examples per task should not be substituted for the quantitative evaluation protocol and held-out scene counts reported in the paper.
  3. Task order, descriptions, and train/evaluation membership follow the curated fly-around render manifest.

Real-world tabletop manipulation

Four real-world task families use teleoperated demonstrations and real-to-sim reconstruction. RIR achieves 75% comprehension with one or three demonstrations. Trajectory success rises from 42% to 57%; three-shot Rw/oR achieves 10%. A separate physical execution case study completes 13 of 20 runs.

  1. The execution case study has one reconstruction failure, two refinement failures, and four missed grasps. Plans currently execute open-loop.

2D reasoning

The corpus contains 25 spatial tasks and 10 spatial-and-algorithmic tasks. With three demonstrations, RIR achieves 58.9% top-1 execution success on spatial tasks versus 22.4% for Rw/oR; corresponding spatial-and-algorithmic results are 24.0% versus 14.5%.

  1. Top-k comprehension and top-1 execution are different metrics. RIR top-5 comprehension approaches human top-1 performance, but RIR top-1 comprehension remains lower. These execution results use noiseless simulated rollouts.

Other formats: Plain text guide · Structured JSON metadata