related_work.md

Related Work and Novelty Check

Date: 2026-04-01

This note is a first-pass literature review for the new research-style additions in this repository. It is intentionally pragmatic: the goal is not to survey everything, but to answer one question clearly.

Question:

• Which parts of this repo still have room for a paper-level novelty claim, and which parts are already well-covered by prior work?

Scope:

• Differentiable / sampling-based MPC
• Neural SDFs for navigation and planning
• GPU-native simulation and RL
• GPU neuroevolution

Papers Checked

Differentiable MPC / MPPI

1. Path Integral Networks: End-to-End Differentiable Optimal Control Link: https://arxiv.org/abs/1706.09597 Date: 2017-06-29 Why it matters:

   • PI-Net already makes the path-integral control computation differentiable end-to-end.
   • This means "differentiable path-integral control" as a broad idea is not new.

2. Differentiable MPC for End-to-end Planning and Control Link: https://arxiv.org/abs/1810.13400 Date: 2018-10-31 Why it matters:

   • This is a canonical differentiable MPC paper.
   • It learns dynamics and cost through the controller, so "differentiate through a planner/controller" is already established.

3. Feedback-MPPI: Fast Sampling-Based MPC via Rollout Differentiation Link: https://arxiv.org/abs/2506.14855 Date: 2025-06-17, revised 2025-12-12 Why it matters:

   • This is the closest recent overlap to our current diff_mppi.
   • It explicitly augments MPPI with sensitivity-analysis-based feedback gains derived from rollout differentiation.

4. GPU Based Path Integral Control with Learned Dynamics Link: https://arxiv.org/abs/1503.00330 Date: 2015-03-01 Why it matters:

   • GPU-accelerated MPPI itself is old.
   • So "MPPI on GPU" alone is not a novelty claim.

Neural SDFs for Navigation

5. iSDF: Real-Time Neural Signed Distance Fields for Robot Perception Link: https://arxiv.org/abs/2204.02296 Date: 2022-04-05 Why it matters:

   • Real-time neural SDF reconstruction for robotics is already established.
   • The paper explicitly highlights downstream collision costs and gradients for planners.

6. Stochastic Implicit Neural Signed Distance Functions for Safe Motion Planning under Sensing Uncertainty Link: https://arxiv.org/abs/2309.16862 Date: 2023-09-28 Why it matters:

   • Neural SDFs have already been used directly inside motion planning under uncertainty.
   • This pushes beyond representation learning into planning-time safety reasoning.

7. Differentiable Composite Neural Signed Distance Fields for Robot Navigation in Dynamic Indoor Environments Link: https://arxiv.org/abs/2502.02664 Date: 2025-02-04, revised 2025-03-06 Why it matters:

   • This is the closest recent overlap to our neural_sdf / sdf_potential_field / sdf_mppi line.
   • It directly targets robot navigation with differentiable neural SDFs and trajectory optimization.

GPU Simulation / RL

8. Isaac Gym: High Performance GPU-Based Physics Simulation For Robot Learning Link: https://arxiv.org/abs/2108.10470 Date: 2021-08-24 Why it matters:

   • GPU-resident simulation plus GPU-resident policy training is already a known research direction.
   • This is the central prior work for anything framed as "MiniIsaacGym".

9. Learning to Walk in Minutes Using Massively Parallel Deep Reinforcement Learning Link: https://arxiv.org/abs/2109.11978 Date: 2021-09-24 Why it matters:

   • Massive parallel RL with thousands of environments on one GPU is already well demonstrated.
   • So "many environments + RL on one GPU" is not novel by itself.

GPU Neuroevolution

10. EvoRL: A GPU-accelerated Framework for Evolutionary Reinforcement Learning Link: https://arxiv.org/abs/2501.15129 Date: 2025-01-25 Why it matters:

    • End-to-end GPU EvoRL frameworks now exist.
    • This weakens any broad novelty claim around "GPU evolutionary RL".

11. TensorNEAT: A GPU-accelerated Library for NeuroEvolution of Augmenting Topologies Link: https://arxiv.org/abs/2504.08339 Date: 2025-04-11 Why it matters:

    • GPU neuroevolution libraries with strong acceleration claims are already current.
    • This especially weakens "GPU neuroevolution framework" as a standalone paper claim.

What Prior Work Already Covers

1. diff_mppi

Prior work already covers:

• differentiable control/planning in general
• path-integral control made differentiable
• rollout-differentiation ideas near MPPI
• GPU MPPI

Implication:

• The phrase "Differentiable MPPI" by itself is too broad to claim as novel.

What may still be claimable:

• a very lightweight hybrid algorithm that keeps vanilla MPPI intact and adds a cheap local autodiff refinement stage in plain CUDA
• a practical compute-quality tradeoff result: for the same control quality, the hybrid method may need fewer samples than vanilla MPPI
• a minimal educational systems contribution: showing how to bolt first-order sensitivity onto a sampling controller without a full differentiable simulator stack

Current novelty score:

• Medium at repo/demo level
• Low to medium at paper level unless we add stronger empirical evidence

2. neural_sdf, sdf_potential_field, sdf_mppi, comparison_sdf_nav

Prior work already covers:

• neural SDF reconstruction
• neural SDF gradients for planning
• safety-aware motion planning with neural SDFs
• recent differentiable composite neural SDF navigation in indoor environments

Implication:

• A 2D learned SDF heatmap plus planner integration is not enough for a paper claim.
• Right now this line is best viewed as a compact demonstration, not a research contribution.

What may still be claimable:

• almost nothing in the current 2D toy setup
• maybe an education/demo paper or tutorial artifact, but not a strong research novelty claim

Current novelty score:

• High as a didactic repo addition
• Low as a research-paper contribution

3. mini_isaac, mini_isaac_rl

Prior work already covers:

• all-GPU simulation loops
• all-GPU policy learning
• thousands of parallel environments
• minute-scale training in massively parallel settings

Implication:

• "MiniIsaacGym" is not a research novelty claim.
• It is useful as a minimal implementation, teaching artifact, or systems reproduction.

What may still be claimable:

• a minimal-from-scratch CUDA teaching implementation
• a compact reproduction-style benchmark

Current novelty score:

• High as an educational engineering artifact
• Very low as a paper contribution

4. neuroevo, comparison_neuroevo

Prior work already covers:

• GPU-accelerated neuroevolution libraries
• GPU-resident evolutionary RL frameworks
• population-scale benchmarking on standard control tasks

Implication:

• "4096 policies evolved on GPU" is not enough for a new paper.

What may still be claimable:

• only if we introduce a genuinely new evolutionary operator, representation, or hardware-aware scaling result

Current novelty score:

• Medium as a repo demo
• Low as a paper contribution

Bottom Line

As of 2026-04-01, the strongest paper candidate in this repo is still diff_mppi.

Why:

• It is the only line where our implementation direction still has some room between prior work and a clean, compact claim.
• It sits at the intersection of: sampling-based control first-order sensitivity lightweight GPU implementation
• The nearest overlap, Feedback-MPPI from 2025, is close enough that we must be precise, but not so close that the direction is dead.

The other new projects are not strong novelty candidates right now:

• neural_sdf_*: good demo, weak paper novelty
• mini_isaac*: good educational systems artifact, weak paper novelty
• neuroevo: useful demo, weak paper novelty after 2025 GPU EvoRL / TensorNEAT papers

If We Want a Paper, the Claim Must Get Narrower

Bad claim:

• "We propose differentiable MPPI."

Why bad:

• too broad
• conflicts with PI-Net, differentiable MPC literature, and especially Feedback-MPPI

Better claim:

• "We study a minimal hybrid controller that combines vanilla MPPI with a local autodiff refinement stage, and show that this refinement improves the quality-vs-samples tradeoff under a fixed GPU budget."

Even better claim:

• "For fixed wall-clock or fixed sample budget, a plain-CUDA MPPI + autodiff refinement controller reaches lower trajectory cost / better goal success than vanilla MPPI across several navigation tasks."

That is a narrower and more defensible systems/control paper.

What Evidence Is Missing for diff_mppi

To move from demo to paper direction, we still need:

1. Controlled baselines

• vanilla MPPI
• our current hybrid diff_mppi
• if possible, a simple feedback baseline or gradient-only ablation

2. Fixed-budget comparisons

• same wall-clock budget
• same number of rollouts
• same horizon
• same GPU

3. Multiple tasks

• easy cluttered navigation
• narrow passage
• higher-speed scenario
• dynamic obstacle variant if feasible

4. Metrics

• success rate
• final goal distance
• average cumulative cost
• collision count
• wall-clock per control step
• rollouts required to reach a given success level

5. Ablations

• no gradient refinement
• one gradient step vs multiple
• autodiff step size / weight
• sampling count sweep

Without these, diff_mppi is still mostly a nice demo.

Recommended Direction

If the goal is publication rather than just a strong repo, I would focus the repo around one paper target:

Title sketch:

• Lightweight Rollout-Differentiated MPPI in Plain CUDA
• Hybrid Sampling-and-Gradient MPPI for Budget-Constrained GPU Control

Core thesis:

• Keep MPPI as the main optimizer.
• Do not replace it with a full differentiable planner.
• Add only a cheap local sensitivity-based refinement step.
• Show better quality per unit compute.

That framing is much stronger than trying to claim novelty for:

• neural SDF navigation
• mini Isaac Gym
• generic GPU neuroevolution

Immediate Next Work

1. Turn diff_mppi into an experiment suite instead of a single demo

• fixed seeds
• CSV logging
• benchmark configs

2. Add a clean ablation binary

• benchmark_diff_mppi
• output success/cost/time tables

3. Add 3 to 5 environments designed for paper figures

• open field
• cluttered maze
• narrow passage
• moving obstacle

4. Write a sharper novelty statement

• explicitly position against PI-Net, Differentiable MPC, and Feedback-MPPI

Honest Assessment

The repo now contains several high-quality research-style demos. But only one line currently looks plausibly paper-worthy with additional work: diff_mppi.

Everything else is better treated as:

• strong engineering
• useful educational artifacts
• supporting material for the repo

Not as the main novelty claim.