Anonymous supplementary material

Reading guide

SPROUT: Generalized Planning Interfaces for Agentic Robotics

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

SPROUT implements a generalized planning interface in which agents combine goals, skills, constraints, and prescribed motion. This anonymous supplement includes the paper, synchronized illustrative tool-use and drawing walkthroughs, an interactive Figure 3 graph-growth animation, success and timing results, recovery case studies, and a rollout gallery.

Suggested reading instructions

  1. Overall success is measured over 24 simulated task–scene instances: eight task families with three scenes each. The reported mean wall-time comparison covers the first five task families only (15 instances).
  2. Selected videos illustrate behaviors and are not the full evaluation set. Videos are sped up as labeled.
  3. The recovery videos show physical retries; they do not measure agent reasoning or per-replan latency.
  4. Synchronized Agent/SPROUT specifications are explanatory schematics, not recorded agent transcripts; phase boundaries follow visible actions at approximate playback times.
  5. The interactive Figure 3 animation is a faithful explanatory graph schematic, not a recorded solver trace. Agent inputs remain fixed; synthesis and search can interleave.

Components in suggested reading order

Anonymous paper

The supplied anonymous draft (version 17), with full method, evaluation, and limitations.

Tool-use specification walkthrough

A tool-keypoint trajectory frees the block, then the agent specifies On(block, plate). The right-side illustration follows video playback and explains which choices SPROUT completes.

Generalized planning interface

Partial specifications become constraint graphs; graph synthesis, sampling and pooled numerical refinement complete the unspecified physical choices.

Figure 3 graph-growth animation

The pen identity and pen path from the agent, together with the initial scene state, seed a graph. Animated stages introduce a shared grasp, kinematic relations, inverse kinematics, and collision-free connecting motion. Sampling and refinement complete the open variables. Includes math-formatted labels, playback controls and independently inspectable stages.

Clear and draw walkthrough

An illustrative Clear(whiteboard) goal precedes a specified pen-tip path. The original video shows clearing the board and drawing two eyes and a curved smile.

Simulation evaluation

SPROUT succeeds on 16/24 instances, compared with 9/24 for Skill+EE and TAMP, and 4/24 for end-effector control. Runtime comparisons must be interpreted with success rates.

Reasoning and planning time

On the first five tasks, SPROUT averages approximately 315 seconds compared with EE 558 seconds; these are subset timing means, including unsuccessful runs. Overall success on the full benchmark is 16/24 for SPROUT and 4/24 for EE. Figure 6 reports 78.4% reasoning and 12.4% planning for SPROUT; its aggregation basis is not separately specified.

  1. Time shares are not absolute durations or per-replan latency. Episode means include unsuccessful runs.

Failure recovery case study

A tool-use attempt leaves a block in a tube. The robot regrips the rod, frees the block, sets the tool aside and places the block on a plate. Also includes banana and utensil placement retries.

  1. Approximate chapter timestamps are playback times, not measured replanning times.

Other formats: Plain text guide · Structured JSON metadata