Generalized Inverse Kinematics via Conditional Flow

Kaitlyn Zhang, Tomás Lozano-Pérez, Leslie Pack KaelblingarXiv 2026 · arXiv

TL;DR

GenGenIK is a generative modeling framework that redefines Inverse Kinematics (IK) as a conditional C-space completion task. By using conditional flow matching, it rapidly produces diverse, collision-free robot configurations that satisfy multiple end-effector targets, mobile base placements, and behavioral preferences—outperforming traditional numerical solvers in speed by up to 34x.

What Problem Does This Paper Address?

In robotics, Inverse Kinematics (IK) is the bridge between a high-level plan (e.g., "grab that mug") and the low-level joint angles needed to execute it. Traditional IK solvers typically treat each query as an isolated problem, returning a single solution.

However, for complex, long-horizon tasks, a single solution isn't enough. A robot might find a valid pose to grab a mug but realize a second later that its arm is now blocking its own camera or that it can't reach the next target without moving its base. Planners need a distribution of valid solutions that account for scene geometry, future reachability, and preferences like stability or manipulability. While numerical solvers can be "nudged" to find multiple solutions, they are often too slow for real-time planning in cluttered environments.

Methods

The authors introduce GenGenIK, which models generalized IK as a conditional generation problem using Masked Conditional Flow Matching (CFM).

The Architecture

At its core, GenGenIK uses a Transformer-based encoder-decoder architecture.

  • Task Encoder: Processes end-effector targets and masks. The masking allows a single model to handle varying numbers of targets and end-effectors.
  • Scene Conditioning: Uses a PointNet++ encoder to process 3D point clouds, allowing the model to "see" obstacles and generate collision-free configurations directly.
  • Solution Decoder: Generates the robot's configuration (base position and joint angles) by transporting a simple Gaussian distribution toward the target distribution of feasible solutions.
Fig. 2. Conditional architecture. Task poses and scene geometry form conditioning tokens, while the current candidate is represented by base and joint tokens. The decoder combines these with FK feedback and optional solution requests.

Hybrid Inference

GenGenIK doesn't just rely on the neural network. It uses a three-stage inference pipeline:

  1. Few-step Sampling: The flow-matching model generates a high-quality initial guess in just a few steps.
  2. Inference-time Guidance: For differentiable constraints (like maximizing manipulability), the model uses gradient-based energy guidance to shift the distribution during sampling.
  3. Refinement: A lightweight numerical optimizer (Levenberg-Marquardt) performs a final "polish" to ensure millimeter-level task accuracy.

Datasets

The models were trained on a massive scale to ensure robustness:

  • Scene-free IK: 10 million configurations per robot.
  • Scene-conditioned Tasks: Approximately 600,000 grasp examples in "MolmoSpaces" (simulated kitchens) and 500,000 examples across 2,500 procedurally generated environments.
  • Evaluation: Tested on held-out scenes including kitchens, bathrooms, and living rooms to check for generalization.

Key Insights and Ideas

  • IK as "Infilling": By treating IK as a masked completion task, the model can perform "warm starts" (completing a pose given a partial configuration) or "trajectory infilling" (finding intermediate poses) without changing the architecture.
  • The Power of the Whole Path: One of the most striking findings is that conditioning the first pose of a trajectory on the entire future path significantly improves success. A valid first pose is 2.4x more likely to be "extendable" to the rest of the task when the model sees the whole plan upfront (Table IV).
  • Unified vs. Specialist: A single unified model trained on multiple task families (lifting, pushing, opening doors) performs just as well as specialized models trained on only one task. This suggests the model learns a general "kinematic grammar" that transfers across behaviors.
  • Scene Awareness: Unlike "scene-agnostic" samplers that generate poses and then check for collisions, GenGenIK's scene-conditioning concentrates its samples in collision-free regions from the start.

Key Results

GenGenIK demonstrates significant improvements over state-of-the-art baselines like cuRobo and IKDiffuser:

  • Speed: On complex path-IK tasks (like bimanual lifting or under-table pickup), GenGenIK reaches a valid solution up to 34x faster than batched numerical optimization.
  • Solve Rate: In a "flip-up" manipulation task, GenGenIK achieved a 98% success rate, compared to only 22% for the baseline numerical solver.
  • Preference Control: The model can be biased at test-time to prefer specific postures (e.g., "elbow up") or base placements without needing retraining.
Preference-conditioned IK allows users to shift the generated distribution toward specific requests, such as base stand-off distances or arm postures.

Limitations and Future Work

While GenGenIK generalizes well to new room layouts, it struggles with qualitatively new scene interactions. For example, if the model was never trained to reach into deep, confined volumes (like the back of a narrow cupboard), it fails to extrapolate that behavior to new scenes.

The authors suggest that future work could involve scaling the training data even further or integrating the model with high-level foundation models to automatically generate the "preference tokens" that guide the robot's style and behavior.

Summary updated

Abstract

Inverse kinematics (IK) is a foundational primitive in robotics, mapping the task-space objectives expressed by higher-level planners to robot configurations that realize them. Traditional IK methods generally treat each query as an isolated pose-to-configuration problem, but planners often need a distribution over valid realizations that accounts for future reachability and preferences such as collision avoidance and stability. We introduce GenGenIK, a generative modeling approach that supports mobile bases, multiple end-effectors and targets, scene awareness, and preferences over robot behavior within a common interface. Experiments show that GenGenIK outperforms existing methods in both solve rate and runtime across tasks such as bimanual lift and flip-up, while allowing flexible control over the sampled distribution and remaining effective under partial scene observations. Website: https://icra-2027-gengenik.github.io/.
BibTeX
@misc{zhang2026generalized,
  author = {Kaitlyn Zhang and Tomás Lozano-Pérez and Leslie Pack Kaelbling},
  title = {{Generalized Inverse Kinematics via Conditional Flow}},
  year = {2026},
  eprint = {2611.15565},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.15565},
}
Figure