Planning and control
Motion planning, trajectory optimization, model-predictive control and task-and-motion planning.
Start here: seminal papers
A Formal Basis for the Heuristic Determination of Minimum Cost Paths★ Must read
IEEE Transactions on Systems Science and Cybernetics, July 1968
The paper defines A*, a graph search that always expands the open node with the smallest f̂ = ĝ + ĥ, and proves it always finds a minimum-cost path when ĥ never overestimates the remaining cost. Under a consistency assumption on ĥ, no admissible algorithm that uses no more information expands fewer nodes.
Analysis of the Computed Torque Drive Method and Comparison with Conventional Position Servo for a Computer-Controlled ManipulatorSeminal
JPL Technical Memorandum 33-601, March 1973
A linear, single-joint analysis of the Stanford arm's computed-torque control (model-based feedforward torque plus inertia-scaled position, rate and integral feedback) against a conventional analog position servo. Computed torque follows the plan with zero error if the model is exact, while the conventional servo, even at a 100 rad/s natural frequency, lags about 1.15 degrees on a 5-second ±50-degree move.
Hybrid Position/Force Control of ManipulatorsSeminal
Journal of Dynamic Systems, Measurement, and Control, June 1981
A controller architecture that splits a task frame's directions into force-controlled and position-controlled ones with a binary selection matrix, while every joint contributes to both. On two axes of a Scheinman arm it tracks 1-10 Nt force steps with about 0.15 s rise time and under 1 Nt steady-state error.
Measure of Manipulability for Robot ManipulatorsSeminal
Journal of the Robotics Society of Japan, February 1984
Study the manipulability measure w = √det(J Jᵀ), proportional to the volume of the end-effector velocity ellipsoid, and use it to find the best working posture of common arm types and a robot finger. A two-link arm is best with its elbow at 90° for any link lengths, and equal links are best for a fixed total length.
Impedance Control: An Approach to Manipulation, Part I: Theory★ Must read
Journal of Dynamic Systems, Measurement, and Control, March 1985
Hogan argues that a manipulator in contact cannot be controlled by commanding position or force alone; it must also be given an impedance, the force it produces in response to motion the environment imposes. Using bond graphs, he shows that a broad class of nonlinear manipulators splits into a virtual trajectory plus an impedance, and that impedances acting on the environment add even when they are nonlinear.
Obstacle Avoidance for Kinematically Redundant Manipulators in Dynamically Varying EnvironmentsSeminal
The International Journal of Robotics Research, September 1985
Keep a redundant arm's end-effector exactly on its commanded trajectory and use the Jacobian's null space to push the link point closest to an obstacle away from it. In simulation a 9-DoF arm spray-painting through a car-door window runs at a 102 ms control cycle.
A Robust Layered Control System for a Mobile RobotSeminal
IEEE Journal on Robotics and Automation, March 1986
Brooks splits mobile-robot control into stacked layers of behaviors (avoid, wander, explore) instead of a perceive-model-plan-act chain, and higher layers take over by suppressing the wires of lower ones. A real robot running the system went a number of hours in a lab and a machine room without hitting a moving or fixed obstacle.
Real-Time Obstacle Avoidance for Manipulators and Mobile RobotsSeminal
The International Journal of Robotics Research, March 1986
Khatib moves collision avoidance out of the path planner and into the low-level controller. The goal attracts the end effector and obstacles repel points on the arm through artificial potential fields, applied as operational-space forces. On a PUMA 560 the servo ran at 225 Hz and avoided moving obstacles seen by a vision system in real time.
A Unified Approach for Motion and Force Control of Robot Manipulators: The Operational Space Formulation★ Must read
IEEE Journal on Robotics and Automation, February 1987
Khatib writes a manipulator's dynamics directly at its end-effector and uses them in one controller for both motion and contact force, extended to redundant arms and kinematic singularities. On a PUMA 560 it ran with a 200 Hz servo loop, and square-wave force commands gave rise times under 0.02 s.
Adaptive Control of Mechanical ManipulatorsSeminal
The International Journal of Robotics Research, June 1987
An adaptive computed-torque controller that estimates unknown mass and friction parameters on-line, with a Lyapunov proof that all signals stay bounded and tracking errors go to zero. On a simulated two-link arm, link masses (3 kg guessed, 4 kg true) and Coulomb friction (0 guessed, 2 Nm true) converge within several seconds.
Numerical Filtering for the Operation of Robotic Manipulators through Kinematically Singular ConfigurationsSeminal
Journal of Robotic Systems, December 1988
Damped least-squares inverse kinematics with a cheap running estimate of the smallest singular value, plus extra damping applied only along the near-singular direction, lets a simulated arm pass through singularities with bounded joint velocity. At a planar arm's internal singularity, peak tracking error falls from 22% to under 4% of the commanded velocity compared with uniformly damped least squares.
Probabilistic Roadmaps for Path Planning in High-Dimensional Configuration Spaces★ Must read
IEEE Transactions on Robotics and Automation, August 1996
The paper builds a roadmap once, from random collision-free configurations joined by a fast local planner, and answers each start-goal query by connecting both ends to it and searching the graph. For 7-dof planar robots among narrow passages, 60 s (fixed base) or 70 s (free base) of learning connects all eight hard test configurations in all 30 roadmaps, and a query then takes a fraction of a second.
Rapidly-Exploring Random Trees: A New Tool for Path Planning★ Must read
Technical Report TR 98-11, Iowa State University, October 1998
A search tree that grows by sampling a random state and steering the nearest vertex a small step toward it, so two states never have to be joined exactly and nonholonomic and kinodynamic systems can be planned for directly. In 2D convex spaces its paths are on average 1.3 to 2.0 times the optimal length, and the report applies it to problems of up to twelve degrees of freedom.
Robot Dynamics: Equations and AlgorithmsSeminal
IEEE International Conference on Robotics and Automation (ICRA), April 2000
A review of rigid-body robot dynamics that writes inverse dynamics (RNEA), the joint-space inertia matrix (CRBA) and forward dynamics (ABA) for tree-structured robots in one spatial-vector notation, then extends them to closed kinematic loops. It traces the move from O(N^4) Lagrangian methods to O(N) recursive algorithms; the ABA needs fewer operations than the CRBA once N > 9.
RRT-Connect: An Efficient Approach to Single-Query Path PlanningSeminal
IEEE International Conference on Robotics and Automation (ICRA), April 2000
RRT-Connect grows two Rapidly-exploring Random Trees, one from the start and one from the goal, and greedily extends each toward the other's newest vertex until they meet. It is probabilistically complete, needs no preprocessing, and solves 2D problems in 0.13-1.52 s and a 3D piano-moving problem in 12.5 s on a 270 MHz workstation.
CHOMP: Gradient Optimization Techniques for Efficient Motion PlanningSeminal
IEEE International Conference on Robotics and Automation (ICRA), May 2009
CHOMP plans by optimizing the trajectory directly, using a covariant gradient step that keeps it smooth while pushing it out of obstacles, starting from a straight line that may pass through obstacles. Without a separate planner it found collision-free paths for 99 of 105 household problems on a WAM arm, and it is the footstep controller for the LittleDog quadruped.
Sampling-based Algorithms for Optimal Motion Planning★ Must read
May 2011
Proves that standard PRM and RRT do not converge to optimal paths, and introduces PRM*, RRG and RRT*, which connect each sample within a radius shrinking like (log n / n)^(1/d) and converge almost surely to the optimum at a constant-factor extra cost. In a cluttered 2D test, RRT averages about 1.5x the optimal cost while every RRT* run converges to it.
Finding Locally Optimal, Collision-Free Trajectories with Sequential Convex OptimizationSeminal
Robotics: Science and Systems (RSS), June 2013
Trajopt plans robot motion by sequential convex optimization, penalizing collisions through signed distances to the volume a link sweeps between waypoints, so paths are safe in continuous time. With four initializations it solved 99% of 198 PR2 arm problems in 0.32 s on average, against 97% in 1.2 s for OMPL's RRTConnect and 85% in 6.0 s for CHOMP.
Fast Marching Tree: a Fast Marching Sampling-Based Method for Optimal Motion Planning in Many DimensionsSeminal
June 2013
FMT* grows a tree of paths outward in cost-to-arrive over a fixed set of samples, collision-checking only the locally best edge, and is proven asymptotically optimal with the first convergence-rate bound for optimal sampling-based planners. On SE(2) benchmarks it reaches a given solution quality about 2x faster than RRT* and 10x faster than PRM*.
Informed RRT*: Optimal Sampling-based Path Planning Focused via Direct Sampling of an Admissible Ellipsoidal HeuristicSeminal
April 2014
Informed RRT* is RRT* that, once it has a path, samples only the ellipsoid of states that could still shorten it, and samples it directly instead of by rejection. In a narrow-gap test it finds the path through the gap in 4.00 s, where RRT* needs 12.32 s.
Batch Informed Trees (BIT*): Sampling-based Optimal Planning via the Heuristically Guided Search of Implicit Random Geometric GraphsSeminal
May 2014
BIT* treats batches of random samples as an implicit graph and searches it in heuristic order like A*, reusing the tree as each batch adds samples. On a hard 14-DOF two-arm problem on HERB it found a solution in 68% of trials within 2.5 minutes, against 36% for FMT* and 8% for Informed RRT*.
Embed to Control: A Locally Linear Latent Dynamics Model for Control from Raw ImagesSeminal
June 2015
E2C trains a variational autoencoder whose latent dynamics are forced to be locally linear, so iLQR or AICO can plan directly on embeddings of raw images. It swings up a pendulum from 48x48 pixels in 90% of trials and balances a cart-pole and controls a three-link arm from images in 93% and 90%, where autoencoder baselines never succeed.
Deep Kalman FiltersSeminal
November 2015
Deep Kalman Filters replace a Kalman filter's linear transition and emission with neural networks and learn them by maximizing a variational lower bound, using an RNN recognition network for inference. On claims data from 8,000 diabetic patients, sampling forward without anti-diabetic drugs gives far more patients with high glucose and A1c than sampling with them.
Value Iteration NetworksSeminal
February 2016
VIN builds value iteration into a policy network as a recurrent convolution, so the policy learns a model to plan on and is trained end-to-end with ordinary imitation learning or RL. On unseen 28x28 grid-worlds it reaches the goal 97% of the time, against 74% for a CNN policy.
Control Barrier Function Based Quadratic Programs for Safety Critical SystemsSeminal
September 2016
Turns barrier-function safety conditions into constraints that are linear in the control input and solves them together with a relaxed control Lyapunov function in a quadratic program, so safety is a hard constraint and performance a soft one. In simulated lane keeping the car stays within 0.9 m of the lane center at no more than 0.3 g lateral acceleration.
Deep Visual Foresight for Planning Robot MotionSeminal
October 2016
Train an action-conditioned video prediction model on 50,000 unlabeled robot pushes, then push objects by choosing actions whose predicted pixel motion carries a user-chosen pixel to its goal. On objects not seen in training it ends 2.52 pixels from the goal on average, against 3.19 for the best calibrated servoing baseline.
Socially Aware Motion Planning with Deep Reinforcement LearningSeminal
March 2017
SA-CADRL trains a deep-RL collision-avoidance value network with a small penalty for breaking a social norm, such as passing on the wrong side, and extends it to four agents. The right-handed policy picks the norm-respecting side in 83-98% of two-agent test cases, at about 20% more extra time to goal. It drove a robot fully autonomously among pedestrians at 1.2 m/s.
Safe Model-based Reinforcement Learning with Stability GuaranteesSeminal
May 2017
Model-based RL that improves a neural network policy using a Gaussian-process dynamics model and only collects data inside a region of attraction certified by a Lyapunov function. On a simulated inverted pendulum, the certified safe region grows from a small initial set over 50 data points and the pendulum never falls.
Information Theoretic Model Predictive Control: Theory and Applications to Autonomous Driving★ Must read
July 2017
A sampling-based MPC derived from information theory averages thousands of GPU-simulated control sequences, weighted by their exponentiated cost, with no gradients and arbitrary nonlinear dynamics. On a 1:5 scale rally car on a dirt track it completed all 100 counter-clockwise laps at an 11 m/s target with a top speed of 9.06 m/s; a cross-entropy MPC baseline completed 66% of its laps at that setting.
Hamilton-Jacobi Reachability: A Brief Overview and Recent AdvancesSeminal
September 2017
A tutorial on Hamilton-Jacobi reachability. It shows how the set of states that can be driven into a target is computed by solving an HJ Isaacs PDE on a grid, introduces the MATLAB and GPU tools that do this, and reviews ways around the exponential cost in state dimension, such as decomposing a 10D quadrotor into small subsystems.
Trajectory Deformations from Physical Human-Robot InteractionSeminal
October 2017
Add an online deformation of the robot's future desired trajectory to impedance control, so a human's push changes where the robot plans to go and not just where it is. In a 10-subject study on a wrist exoskeleton, users applied about half the torque (2.25 vs 4.64 Nm) and moved more smoothly than with impedance control alone.
Riemannian Motion PoliciesSeminal
January 2018
RMPs pair each motion policy with a Riemannian metric, so controllers written in many task spaces can be pulled into joint space and combined as a metric-weighted average that is provably optimal. With up to 150 controllers on a 7-DoF arm, RMPs solved every cluttered reaching task, while combining the same policies without their metrics led to periodic collisions.
PDDLStream: Integrating Symbolic Planners and Blackbox Samplers via Optimistic Adaptive PlanningSeminal
February 2018
PDDLStream extends the PDDL planning language with streams, blackbox samplers such as IK solvers and motion planners, and solves the result as a sequence of finite PDDL problems. Its new Adaptive algorithm solves about 98% of 5-block packing problems, where the earlier algorithms solve about 10%.
World ModelsSeminal
March 2018
A VAE compresses each frame, an MDN-RNN predicts the next latent, and a linear controller with under 1,100 parameters is evolved on top. It scores 906 ± 21 on CarRacing-v0, and a VizDoom policy trained only inside the learned model survives 1092 ± 556 steps in the real game (750 counts as solved).
Deep Reinforcement Learning in a Handful of Trials using Probabilistic Dynamics ModelsSeminal
May 2018
PETS models the dynamics with an ensemble of probabilistic neural networks and plans through it with sampled particle trajectories and MPC. It reaches model-free asymptotic performance on four MuJoCo tasks in under 100 trials, with 8x fewer samples than SAC and 125x fewer than PPO on half-cheetah.
Motion Planning NetworksSeminal
June 2018
MPNet learns motion planning from RRT* demonstrations. One network encodes the obstacle point cloud and another predicts the next configuration. In 2D, 3D and rigid-body tests it finds paths in 0.11-0.48 s on average, 18-36x faster than BIT*, including in workspaces it never saw.
Crowd-Robot Interaction: Crowd-aware Robot Navigation with Attention-based Deep Reinforcement LearningSeminal
September 2018
A deep RL value network for crowd navigation that encodes every robot-human pair, adds a local map of each human's neighbors, and pools the pairs with self-attention. In simulated crowds that ignore the robot it reaches the goal in all 500 test cases, versus 0.95 for an LSTM baseline and 0.78 for CADRL.
Differentiable MPC for End-to-end Planning and ControlSeminal
October 2018
Make box-constrained MPC a differentiable policy layer by differentiating the convex approximation at the iLQR solver's fixed point, so the backward pass is a single extra LQR solve instead of an unrolled chain. Learning dynamics from expert actions alone matches system identification on pendulum and cartpole and beats it when the expert lies outside the model class.
Learning Latent Dynamics for Planning from PixelsSeminal
November 2018
PlaNet learns a latent dynamics model from 64x64 images and picks every action by planning in that latent space, with no policy or value network. On six DeepMind control suite tasks it comes close to D4PG's final scores, and beats D4PG by 26% on cheetah, while using 1,000 episodes instead of 100,000.
Control Barrier Functions: Theory and Applications★ Must read
March 2019
A tutorial overview of control barrier functions, a Lyapunov-like condition that keeps a system inside a safe set and that a quadratic program can enforce by changing a nominal controller as little as possible. Applications include walking on stepping stones in simulation, lane keeping on small robots, multi-robot coverage, and a Segway safety filter that runs onboard in 0.4 ms on average.
Variable Impedance Control in End-Effector Space: An Action Space for Reinforcement Learning in Contact-Rich TasksSeminal
June 2019
An RL policy outputs both end-effector pose targets and the stiffness of an impedance controller (VICES), compared against joint-space and fixed-impedance action spaces on path following, door opening and wiping. VICES earns the highest wiping reward at a 28 N mean force, transfers from a simulated Panda to a Sawyer, and wipes a real whiteboard in 8 of 10 trials.
Deep Lagrangian Networks: Using Physics as Model Prior for Deep LearningSeminal
July 2019
DeLaN builds the Euler-Lagrange equation into a deep network that learns a robot's inertia matrix and gravity from joint positions, and learns it online during real-time control. Trained on 8 of 20 characters on a simulated 2-dof robot, it tracks all 20 about as well as the analytic model, and it handles faster trajectories much better than a standard network.
Crocoddyl: An Efficient and Versatile Framework for Multi-Contact Optimal ControlSeminal
September 2019
An open-source optimal control library for legged robots with a given contact sequence, built around FDDP, a DDP variant that starts from infeasible guesses and keeps dynamics gaps open in early iterations. Gaits, jumps and a front-flip converge in 10-34 iterations and under 0.5 s, with single iterations running at up to 859.6 Hz.
Highly Dynamic Quadruped Locomotion via Whole-Body Impulse Control and Model Predictive ControlSeminal
September 2019
A convex MPC plans ground reaction forces on a simple lumped-mass model, and a 500 Hz whole-body impulse controller tracks those forces, rather than a body trajectory, using full-body dynamics. On the Mini-Cheetah it runs five gaits outdoors with one set of gains and reaches 3.7 m/s on a treadmill, up from 2.45 m/s with MPC alone.
Dream to Control: Learning Behaviors by Latent ImaginationSeminal
December 2019
Dreamer learns a latent world model from pixels and trains an actor and a value model purely on imagined rollouts, backpropagating value gradients through the learned dynamics. On 20 DeepMind Control tasks it averages 823 after 5 million steps, above D4PG's 786 after 100 million.
Trajectron++: Dynamically-Feasible Trajectory Forecasting With Heterogeneous Data★ Must read
January 2020
A graph-structured generative model that forecasts many interacting agents by predicting control actions and integrating their dynamics, using semantic maps and optionally the robot's own plan. On ETH/UCY pedestrians its best-of-20 errors are 55-60% lower than the prior best.
Integrated Task and Motion Planning★ Must read
October 2020
A review that frames task and motion planning (TAMP) as a search over discrete plan skeletons plus the continuous constraints each skeleton induces. It sorts 33 lines of prior work by whether they sequence actions first, satisfy constraints first or interleave the two, and by whether they use pre-discretized values, sampling or optimization.
Vision-Only Robot Navigation in a Neural Radiance WorldSeminal
October 2021
A robot plans through a pre-trained NeRF by minimizing NeRF density along a differentially flat trajectory and localizes by matching its RGB images to NeRF renders, replanning in a loop. In simulated scenes a quadrotor reaches the goal without collision where open-loop execution crashes, but each 0.1 s step still takes about 2 s to replan and 4 s to estimate state.
IKFlow: Generating Diverse Inverse Kinematics SolutionsSeminal
November 2021
IKFlow trains a conditional normalizing flow per robot that samples many different inverse-kinematics solutions for a goal pose at once, instead of the single answer a numerical solver gives. It returns 500 solutions in 5 ms, with average errors of 0.36-7.72 mm across ten kinematic chains.
Language Models as Zero-Shot Planners: Extracting Actionable Knowledge for Embodied AgentsSeminal
January 2022
Frozen GPT-3 and Codex models write household task plans one step at a time, and a sentence-embedding model maps each step to the closest action VirtualHome can execute. No model is trained, and executability rises from 18% to 79% for Codex 12B, at some cost in human-judged correctness.
Do As I Can, Not As I Say: Grounding Language in Robotic Affordances★ Must read
April 2022
SayCan picks each next step for a mobile manipulator by multiplying a language model's score for every skill description with that skill's learned value function, so the step is both useful and feasible. With PaLM it correctly plans 84% and executes 74% of 101 kitchen instructions.
Motion Planning around Obstacles with Convex OptimizationSeminal
May 2022
GCS plans collision-free trajectories as a shortest path through a graph of convex safe regions, solved as one convex program plus a cheap rounding step. On 100 random quadrotor environments 95% of its plans are within 1% of the global optimum, and on a 7-DoF arm it finds shorter paths than PRM in less time.
Planning with Diffusion for Flexible Behavior SynthesisSeminal
May 2022
Diffuser trains a diffusion model over whole state-action trajectories and plans by denoising them, steered by a learned return gradient or by inpainting the start and goal. On sparse-reward Maze2D it averages a score of 119.5, against 47.0 for the best model-free baseline (IQL).
DayDreamer: World Models for Physical Robot LearningSeminal
June 2022
Runs the Dreamer world-model RL algorithm online on four real robots, with no simulator and one set of hyperparameters. A Unitree A1 learns to roll over, stand up and walk in about 1 hour, and two arms learn visual pick-and-place from sparse rewards at close to human pick rates.
Inner Monologue: Embodied Reasoning through Planning with Language ModelsSeminal
July 2022
A frozen LLM plans robot skills step by step while success detection, scene descriptions and human answers are fed back into its prompt as text, so it can retry and replan. On a real kitchen mobile manipulator this raises success from 30.8% (SayCan) to 60.4%.
Perceptive Locomotion through Nonlinear Model Predictive ControlSeminal
August 2022
A nonlinear MPC for ANYmal that optimizes torso and all leg joints together, with the terrain given as convex foothold regions and a signed distance field for collisions. It crosses gaps, stepping stones and stairs on hardware with a mean MPC iteration of 6.6 ms.
ProgPrompt: Generating Situated Robot Task Plans using Large Language ModelsSeminal
September 2022
Prompt an LLM with a Python-style header (the robot's actions as imports, the scene's objects as a list, a few example task programs) and have it write the new task plan as code with comments and precondition assertions. With GPT-3 it reaches 0.34 success rate on VirtualHome tasks where a prior LLM planner gets 0, and it runs on a real Franka arm for tabletop pick-and-place.
Motion Policy NetworksSeminal
October 2022
MπNets is a neural policy that drives a 7-DoF arm to a task-space target while avoiding obstacles, using only a segmented point cloud. It is trained on 3.27 million expert trajectories and succeeds on 95.33% of test problems, against 65.28% for the earlier neural planner MPNets; each solution takes 0.33 s, versus 16.46 s for a global planner.
VoxPoser: Composable 3D Value Maps for Robotic Manipulation with Language ModelsSeminal
July 2023
GPT-4 writes code that calls open-vocabulary perception to build 3D voxel maps of where the gripper should go and what it should avoid, and a motion planner follows them in closed loop with no training. On five real-robot tasks it succeeds 88% of the time (70% under disturbances), against 24% (0%) for an LLM calling motion primitives.
The Safety Filter: A Unified View of Safety-Critical Control in Autonomous SystemsSeminal
September 2023
A review that casts HJ reachability, control barrier functions, model predictive shielding and learned safety critics as one structure, a fallback policy plus a safety monitor plus an intervention scheme. One Universal Safety Filter Theorem then rederives the safety guarantee of each class.
DTC: Deep Tracking ControlSeminal
September 2023
DTC trains an RL policy in simulation to track footholds from a trajectory optimizer, giving the ANYmal quadruped model-based foot-placement accuracy together with RL robustness. It tracks footholds to 2.3 cm and crosses gaps, beams and stepping stones where an RL baseline fails.
cuRobo: Parallelized Collision-Free Minimum-Jerk Robot Motion GenerationSeminal
October 2023
A GPU library that plans collision-free, minimum-jerk arm trajectories by running many IK, geometric-planning and trajectory-optimization seeds in parallel. On 2600 Franka benchmark problems it plans in 50 ms on average, 60x faster than Tesseract, and solves 99.8% of them.
Splat-Nav: Safe Real-Time Robot Navigation in Gaussian Splatting MapsSeminal
March 2024
Plans and localizes a drone directly in a Gaussian Splatting map. Polytope corridors built from the splat's ellipsoids give provably safe Bézier trajectories, and RGB-only PnP against splat renders gives the pose. Re-planning runs at over 2 Hz and localization at about 25 Hz, and only one of 120 hardware flights collided.
Recent
cuRoboV2-Fast: Fast Torque-Limited Collision-Free Motion Planning
November 2026
A GPU motion planner that optimizes B-spline trajectories with the payload in the dynamics, then retimes them under velocity, acceleration, jerk and torque limits. It plans a 3 kg-payload motion in 4.4 ms on an RTX 5090, 14.7x faster than cuRoboV2, and cuts real pick-and-drop cycle time by 34.6%.
FEIJOA: Task-Driven Grounding and Execution for Mobile Manipulation
November 2026
A language-model agent for mobile manipulation that grounds each instruction against a raw RGB-D scan when it arrives, and checks every skill's feasibility before running it and its effect afterwards. It resolves 94.2% of 103 open-ended grounding queries exactly (68.0% for the best prebuilt map) and completes 35 of 50 real tasks on Spot.
TAMPAgent: Agentic Belief-Space Planning for Mobile Manipulation
November 2026
A language-model agent writes belief-space task-and-motion plans, and perception, grasp and motion tools must certify each action's constraints before the robot moves. On simulated kitchen tasks with an RB-Y1 mobile manipulator it reaches 77.2% success, against 57.2% for Code as Policies and 61.1% for Graph as Policy.
New papers stay for 14 days.
Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.