Efficient Foundation Models
for Real-Time Embodied AI

Making robot foundation models fast enough, cheap enough, and safe enough to actually deploy.

CoRL 2026 · November 12, 2026 · Austin, Texas

Overview

Embodied foundation models are now good enough to matter. Most of them still miss the latency, memory, and power budgets required to run on the robot itself. This workshop is about that gap: what it actually takes to make an embodied model general, fast, and deployable on real hardware rather than in a datacenter.

This is for you if you work on compression, quantization, distillation, adaptive compute, action chunking, hardware-aware design, edge-cloud systems, or safety and control under timing constraints, whether you build vision-language-action policies, world-action models, or something that does not fit either label. Submissions are non-archival and close October 12.

Why now

RT-2's 55B parameters run at 1 to 3 Hz on cloud TPUs that cost more than the robot. OpenVLA reports roughly 6 Hz on an RTX 4090, and that is a desktop GPU rather than anything that rides onboard. Dexterous manipulation generally wants closed-loop rates well above what these policies deliver, and while some systems close that gap with a fast low-level servo loop under a slower policy, the policy rate still bounds how quickly the robot can change its mind. Meanwhile humanoid programs at scale (Tesla Optimus, Agility Digit, Unitree G1) make per-unit inference cost a first-order constraint: on platforms where the robot costs less than a datacenter GPU, onboard compute is a budget rather than a preference.

Model	Params	Hardware	Reported figure	Measures
Action policies
RT-2	55B	Cloud TPU	1–3 Hz	closed-loop control
OpenVLA	7B	RTX 4090	~6 Hz	throughput
π₀	3.3B	—	50 Hz	open-loop chunk rate
TinyVLA	0.7–1.3B	RTX A6000	20× vs OpenVLA	relative latency
OpenVLA-OFT	7B	A100	109.7 Hz (LIBERO)	throughput
MiniVLA	1B	L40S	12.5 Hz	throughput
HyperVLA	90× fewer active	L4	482→4 ms/step	relative latency
BitVLA	1.58-bit ternary	—	11× vs OFT	relative memory
World models
DreamDojo	2–14B	RTX 5090	10.8 FPS	offline rollout
DreamZero	14B	GB200	7 Hz	closed-loop control

The last column is the point: these numbers measure different things on different hardware, so they cannot be ranked against each other. Building an evaluation that can is one of the things we want papers about.

General

One policy across tasks, objects, and embodiments, rather than a model per skill.

Fast

Closed-loop rates that keep up with contact-rich manipulation.

Deployable

Inside a fixed onboard power, memory, and cost budget, not merely running somewhere.

Systems trade these off differently, and few manage all three at once. Speed and deployability pull apart in particular, because raw Hz can be bought with bigger hardware: OpenVLA-OFT reaches 109.7 Hz, but on an A100 that will never ride on the robot.

What we want papers on

Architecture and compression

Smaller backbones, action chunking, quantization, pruning, distillation, and adaptive compute for policies and world models alike. Results reporting real latency, memory, energy, or control-rate gains on robot-relevant hardware are especially welcome, as is any evidence about whether these techniques compose.

Edge-cloud co-design

Deployment trade-offs across onboard, hybrid, and cloud setups, and across heterogeneous accelerators: GPUs, TPUs, Trainium and Inferentia, mobile NPUs. Total cost of ownership at fleet scale is close to unstudied.

Safety under latency

Interruption, fallback, and certification when inference is slow or its timing is variable. Action chunking commits a robot to open-loop execution with no principled way to interrupt it, and certifying systems with stochastic response time remains open.

Speakers

Jiajun Wu

Stanford University

World models for embodied AI: efficient learned models for online planning.

Yecheng “Jason” Ma

Dyna Robotics

Embodied foundation models on a commercial duty cycle, from DYNA-1 to the DYNA-2 world-action model.

Dhruv Shah

Princeton University and Google DeepMind

Post-training as the stage where most deployable-efficiency decisions are made.

Ouais Alsharif

Google DeepMind

Engineering robot foundation models for production deployment.

Additional invited speakers will be announced here as invitations are confirmed.

Schedule

Roughly four hours, built around four interactive elements: live profiling on real Jetson hardware, a structured debate, the challenge results, and a roadmap working session. Talk segments are kept short. Slots marked to be announced are still being filled.

Time	Session
9:00	Opening and framing Organizers frame the generality, speed, and deployability trilemma
9:10	Keynote, speaker to be announced 30 minute keynote, then 20 minutes live: organizers profile vanilla OpenVLA on a Jetson on stage, narrating memory and kernel bottlenecks
10:00	Lightning talks Five contributed papers, selected for method diversity
10:30	Coffee and posters All accepted papers; live demos welcome
10:55	Invited talk, speaker to be announced C3: certifiable safety under latency
11:25	Debate: the 1B ceiling, participants to be announced Oxford-style and moderated, with an audience vote before and after
12:10	Challenge results Top entries present, seven minutes each
12:35	Synthesis: roadmap working session Organizer-facilitated, feeding the post-workshop report
12:55	Best paper and closing

Call for Papers

We solicit four-page extended abstracts, plus unlimited pages for references, on efficiency across the whole space of embodied foundation models: VLA policies, world-action models, and emerging architectures alike.

Topics

Efficient architectures for VLA policies, world-action models, and emerging designs: action chunking, mixture-of-experts, early exit, diffusion and flow step-reduction
Compression for robot foundation models: quantization, pruning, distillation
Memory-efficient adaptation and fine-tuning
Benchmarking methodology, including whether efficiency techniques compose
Hardware-aware design across accelerators: GPU, TPU, AWS Trainium and Inferentia, mobile NPUs
Edge-cloud hybrid architectures for robot fleets
Safety and latency trade-offs, and fallback policies under stochastic response time
Efficient world models for online planning and real-time safety monitoring

Submission

Four pages plus unlimited references, hosted on OpenReview, reviewed double-blind. Submissions must be anonymized.

Non-archival. Accepted papers are not archived as formal publications, so submitting here does not affect your ability to submit the same work to a main conference. Concurrent submissions and work under review elsewhere are permitted. Previously published work that fits the scope is also eligible, tracked and reviewed separately from new work.

Submit at OpenReview. The venue is open now.

OpenReview moderates newly created profiles: a profile made with an institutional email address is activated automatically, while one made without goes through a moderation step. If you are creating an account to submit, please allow time for that.

Review and presentation

We aim for three reviews per submission, from the organizing committee and an invited program committee, subject to reviewer availability. All accepted papers are presented as posters, and roughly five are additionally selected for a five-minute lightning talk.

Embodied Efficiency Challenge

A benchmarking competition producing a concrete, citable artifact. The task is to compress a reference policy until it fits a real onboard budget without losing the capability that made it worth deploying.

Reference model: OpenVLA-7B with a frozen evaluation harness. About 14 GB in bf16, which does not fit a 16 GB Jetson, so quantization, sharding, or offload is required.
Target budget: Jetson Orin NX (16 GB) class hardware. Teams run on their own device.
How it works: We release a reference Docker container and evaluation script. You run it on your hardware and submit code, reported metrics, and a hardware attestation. Organizers re-run the top entries to verify before results are announced, so the leaderboard is reproducibility-checked rather than centrally hosted.
Tasks: LIBERO Spatial, Object, and Goal suites.
Metrics: Success rate first, within 15 points of the published reference. Then median and p99 latency per action. Then parameters and peak memory.
Open to all: The compress-to-a-budget methodology is architecture-agnostic. World-model and emerging-architecture entries are explicitly welcome.
Artifact: All submissions, baselines, and the harness become embodied-efficiency-bench, a permanent open-source resource.
Recognition: Best Efficiency Award and a presentation slot. Travel grant subject to sponsorship.