No Plan but Everything under Control: Robustly Solving Sequential Tasks with Dynamically Composed Gradient Descent

Authors: Vito Mengers, Oliver Brock · ICRA 2025

Licences

Open Access
Paper: CC BY 4.0

Everything below draws on the full paper under its CC BY 4.0 licence; its figure is credited to the authors.

TL;DR

Following the steepest gradient through a network of interconnected estimators solves Blocks World and real drawer opening without a planner, adapting to uncertainty and disturbances.

Why it matters

The method completes multi-step manipulation without computing a plan, and behaviours such as moving the camera to reduce uncertainty, re-finding a lost handle and re-grasping arise from gradient selection rather than being programmed. In the drawer experiments it kept working when the cabinet was moved or the handle was pulled out of the hand, where the planning baselines lost performance.

Problem

Planners predict future states to sequence subgoals, but uncertainty compounds over the planning horizon, so costly plans often become infeasible in dynamic or uncertain settings. Myopic gradient-based control adapts quickly to feedback, yet it has been limited to continuous, single-step problems because subgoals are hard to encode in one potential function that covers all task and environment variations.

Proposed solution

The authors encode the world's regularities in AICON, a network of recursive estimators (components) joined by active interconnections whose information exchange changes with the system state. A differentiable goal cost is expanded with the chain rule along every path to the actuation signals, and at each time step only the steepest of these gradients is followed, so gradients that would change an interconnection's regime act as implicit subgoals. The pattern is applied to Blocks World and to a partially observable drawer-opening task on a Panda arm with an RBO Hand 3, a wrist force-torque sensor and an in-hand RGB camera.

Key insights

Subgoals do not have to be listed: when an active interconnection sits in a stationary regime, the chain rule yields a gradient that tries to change that regime, and this acts as an implicit subgoal.
Following only the steepest gradient avoids conflicting gradients; summing the gradients instead gave unstable, jerky behaviour and occasional failures in the ablation.
Interactive perception (triangulating the drawer by moving the wrist camera) and error recovery emerged from gradient descent without being implemented explicitly.
Tasks with competing subgoals, irreversible actions or required temporary setbacks, such as Towers of Hanoi, still need planning.

Main results

Blocks World (10 to 30 blocks, 0 to 15 towers): both the basic gradient descent and the variant with an actively interconnected goal solved every instance, including ones needing at least 35 steps, close to the optimal move count. The text says over 100 instances were generated; the Fig. 2 caption says 130 randomly generated problems.
Unstacking problems were solved optimally; stacking problems showed extra moves from competing subgoals, reduced but not removed by interconnecting the goal with the current state.
Real drawer opening under varied prior uncertainty, added sensor noise, and light (cabinet moved) or heavy (drawer pulled from the hand) disturbances, ten runs per scenario: 3 failures in seventy trials, all from unmodelled singularities and self-collisions. The abstract counts one hundred real-world drawer trials in total.
The method outperformed a state-space planner with a precomputed trajectory and a belief-space planner with exploratory motions; the belief-space planner handled uncertainty but struggled with large disturbances, and the method was computationally cheaper than the 67-dimensional belief-space planning.
Ablations over three initial configurations and three alternative tasks (n = 5 per setting): reducing active interconnections caused sharp performance drops, most in tasks that needed viewpoint adjustment or grasping.

Limitations

Competing subgoals and missing foresight lead to suboptimal moves; consistently finding Blocks World shortcuts (moving blocks directly between towers) needs planning (stated by the authors).
Tasks with irreversible actions, such as destroying objects, or conflicting subgoals that require temporary setbacks, such as Towers of Hanoi, still require planning (stated by the authors).
All three real-world drawer failures came from effects the system does not model: singularities and self-collisions.
Each task uses its own specified set of estimators, interconnections and goal cost; the paper builds one such system for Blocks World and one for drawer opening.
The real-robot evaluation covers one task family, drawer opening, with ten trials per scenario and five per ablation setting.

Field impact

Won the IEEE ICRA 2025 Best Paper Award on Planning and Control and was a finalist for the ICRA 2025 Best Paper Award on Robot Learning; 10 citations. The paper links a project page with supplementary videos, which links the AICON code repository and tutorials.

Abstract

We introduce a novel gradient-based approach for solving sequential tasks by dynamically adjusting the underlying myopic potential field in response to feedback and the world’s regularities. This adjustment implicitly considers subgoals encoded in these regularities, enabling the solution of long sequential tasks, as demonstrated by solving the traditional planning domain of Blocks World—without any planning. Unlike conventional planning methods, our feedback-driven approach adapts to uncertain and dynamic environments, as demonstrated by one hundred real-world trials involving drawer manipulation. These experiments highlight the robustness of our method compared to planning and show how interactive perception and error recovery naturally emerge from gradient descent without explicitly implementing them. This offers a computationally efficient alternative to planning for a variety of sequential tasks, while aligning with observations on biological problem-solving strategies.

Open Source Robotics