←
gokul@NeurIPS2026 ✈️226 posts
gokul@NeurIPS2026 ✈️ avatar

gokul@NeurIPS2026 ✈️

@gokulp01

PhD candidate @UofIllinois (UIUC) @CSL_Illinois | Current: robotics research intern @Nvidia | ex-research science intern @Adobe Research

◷ Joined September 2017▧ 32 Photos and videos
gokul@NeurIPS2026 ✈️ avatar
gokul@NeurIPS2026 ✈️@gokulp01·Jul 16

Muninn won the Outstanding Student Paper Award at #RSS 2026!! Huge thank you to my collaborators and advisors, this work would not exist without them! And to the #RSS community for the recognition and for organizing such an awesome conference! We have also open-sourced the core repository. The plan is to actively maintain it, and to benchmark and speed up as many diffusion-based trajectory planners as we can get our hands on. If your robot plans with diffusion, try it out! Issues and PRs very welcome. 📄 Paper: arxiv.org/pdf/2605.09999 💻 Code: github.com/gokulp01/Muninn cc: @RoboticsSciSys

We are presenting Muninn at RSS 2026, and we're incredibly honored that it has been selected as a finalist for both the Outstanding Paper & Outstanding Student Paper awards! #RSS2026 TL;DR: diffusion planners generate great robot trajectories, but nobody wants 100 denoiser calls inside a control loop. Existing speedups (truncation, distillation, fixed skipping) trade quality for speed in ways you can't predict before deployment. Muninn is a training-free wrapper that skips denoiser calls only when it can certify the final plan stays within your tolerance 🧵 (1/9) 📄 Paper: arxiv.org/abs/2605.09999 💻 Code: github.com/gokulp01/Muninn

0:15
gokul@NeurIPS2026 ✈️ avatar
gokul@NeurIPS2026 ✈️@gokulp01·Oct 15

Super insightful read. The ablations are the real gem here. The harness matters way more than I expected, and the right one makes action space and prompt details way less of a big deal. + RoboLab is the best sim eval out there rn 🤖

Can and should a general-purpose model serve as a robot policy? In Astranomy, we study GPT-6 Astra for general pick-and-place manipulation. Astra sets a new SOTA on RoboLab-120, but real-time robot control remains an open challenge. Full tech report: research.nvidia.com/labs/srl…

2:07
gokul@NeurIPS2026 ✈️ avatar
gokul@NeurIPS2026 ✈️@gokulp01·Oct 9

📄 Paper: arxiv.org/abs/2610.12386 🌐 Website: arc-robot-reasoning.github.i… 💻 Code: [coming soon] 🤗 Models: [coming soon] 🗂️ Dataset: [coming soon] This work was done during my internship at @nvidia @NVIDIAAI @NVIDIARobotics HUGE thanks to all my collaborators and mentors: Tao Sun, @EliJalbout, @moritz_reuss, @mli0603, Fabio Ramos, @xuningy, @imankitgoyal! More videos below 👇

0:25
0:25
0:25
0:25
gokul@NeurIPS2026 ✈️ avatar
gokul@NeurIPS2026 ✈️@gokulp01·Oct 9

(7/7) Key findings beyond task performance: → Reasoning improves execution across long-horizon, dynamic, and reasoning-intensive tasks. → Nearly invariant to instruction specificity: π0.5+ARC holds ~45% from specific to vague instructions, while its base falls to 15.2%. → Inference efficiency: 10× fewer denoising steps for π0.5 (within 0.4 pp of its base) and 2× fewer for Cosmos3-Nano-Policy (+4.8 pp). → Training efficiency: Cosmos3-Nano-Policy+ARC matches its baseline's 10K-iteration performance with ~4.3× fewer updates.

Post attachment
gokul@NeurIPS2026 ✈️ avatar
gokul@NeurIPS2026 ✈️@gokulp01·Oct 9

(6/7) Stronger grounding of causal language leads to stronger performance. Zero-shot gains over each base model: RoboLab-Reasoning-50 +50.0 pp (Cosmos3-Nano-Policy), +46.8 pp (π0.5); MolmoSpaces +27.2 pp (π0.5), +18.6 pp (Cosmos3-Nano-Policy); RoboLab-120 up to +29.8 pp.

Post attachment
gokul@NeurIPS2026 ✈️ avatar
gokul@NeurIPS2026 ✈️@gokulp01·Oct 9

(5/7) 3️⃣ The training. We show how state-of-the-art VLAs (π0.5) and WAMs (Cosmos3-Nano-Policy) learn to use these traces for control, with fine-tuning and inference tailored to each architecture. No redesign, no new pretraining. At deployment, an external VLM writes the traces asynchronously.

gokul@NeurIPS2026 ✈️ avatar
gokul@NeurIPS2026 ✈️@gokulp01·Oct 9

(4/7) 2️⃣ The data. We introduce ARC-Trace, a pipeline that generates these traces automatically from existing demonstrations, aligned to action chunks. Relabeling DROID gives ARC-Trace-DROID: ~75K episodes and ~1.2M action-aligned frames, without collecting a single new robot demonstration.

1:04
gokul@NeurIPS2026 ✈️ avatar
gokul@NeurIPS2026 ✈️@gokulp01·Oct 9

(3/7) 1️⃣ The reasoning representation. We introduce action-grounded causal traces: reasoning grounded in the robot's next action, explaining its causal structure: why the action is appropriate and what effect it should produce. Unlike subtasks or CoT, policies actually use it: 41% action-head attention vs 4-6%.

Post attachment
gokul@NeurIPS2026 ✈️ avatar
gokul@NeurIPS2026 ✈️@gokulp01·Oct 9

(2/7) ARC has three key ingredients: 1️⃣ The reasoning representation: action-grounded causal traces 2️⃣ The data: a scalable pipeline that labels existing demonstrations with these traces automatically 3️⃣ The training: a recipe that adapts pretrained VLAs and WAMs to use these traces for control

Post attachment
gokul@NeurIPS2026 ✈️ avatar
gokul@NeurIPS2026 ✈️@gokulp01·Oct 9

Today we are introducing ARC, a reasoning recipe for robot foundation models that sets a new state of the art on RoboLab-120 and MolmoSpaces. The prevailing way to improve robot foundation models is brute force: larger models, more robot demonstrations, costly training at scale. ARC needs none of that. With no new robot data and no foundation-scale training, π0.5 and Cosmos3-Nano-Policy gain up to: 📈 +50.0 pp on RoboLab-Reasoning-50 📈 +29.8 pp on RoboLab-120 📈 +27.2 pp on MolmoSpaces To our knowledge, these gains are unprecedented. 📄 Paper: arxiv.org/abs/2610.12386 🌐 Website: arc-robot-reasoning.github.i… 🧵 (1/7)

1:09
gokul@NeurIPS2026 ✈️ avatar
gokul@NeurIPS2026 ✈️@gokulp01·Sep 7

Takeaway: command-preserving trajectory redirection exposes a new failure mode in VLAs: tiny semantic-preserving perturbations can compound through closed-loop control and steer the robot toward a different, adversary-chosen task. (9/9)

gokul@NeurIPS2026 ✈️ avatar
gokul@NeurIPS2026 ✈️@gokulp01·Sep 7

To separate failure from redirection, we compare each attacked action with what the VLA would do under the original and target commands. The attack follows the target, even when the tasks demand different moves. Causal tracing points to the corrupted destination word. (7/9)

Post attachment
gokul@NeurIPS2026 ✈️ avatar
gokul@NeurIPS2026 ✈️@gokulp01·Sep 7

Across 9 VLA families and four LIBERO suites, attack success ranges from 82.8–97.5%; 7 of 9 models exceed 90%. Successful prompts differ from the original by only 3.4 characters on average. This is a shared failure mode across VLA architectures. Not a quirk of a single model. (6/9)

Post attachment