DC-QFA Deploys Robot AI to Any Device Without Retraining, 3x Faster
A single supernet trained once, then specialized per robot, cuts visuomotor policy latency 2-3x with negligible accuracy loss on real hardware.
ProTakeaways
- DC-QFA trains one quantization-aware supernet and specializes it per device via lightweight search.
- Delivers 2-3x speedup and up to 4x memory reduction on Jetson Orin NX with negligible task-success loss.
- Uses NSGA-II multi-objective search over architecture and mixed-precision bit-widths, guided by real-device latency and memory lookup tables.
- Multi-step on-policy distillation aligns the quantized student with a full-precision teacher to fix long-horizon drift.
- Validated on DiffusionPolicy-T, MDT-V, and OpenVLA-OFT across Push-T, CALVIN, and LIBERO benchmarks.
- Real-world Inovo robot with force/torque sensor: W4A4 policies retain contact-rich manipulation success far above PTQ baselines.
Shipping a state-of-the-art visuomotor policy is easy on a lab workstation and painful everywhere else. A Jetson Orin NX has different memory, latency, and precision constraints than a consumer GPU, which behaves nothing like a cloud instance. Adapting a Diffusion Policy or an OpenVLA-style model to each of those targets usually means a fresh round of quantization-aware training or pruning per device, which is slow and produces one-off checkpoints that do not transfer.
A new framework called DC-QFA (Device-Conditioned Quantization-For-All) tries to kill that per-device retraining loop. It trains a single supernet spanning a rich design space over network architectures and mixed-precision bit-widths, using device-conditioned quantization-aware training and hardware-constrained architecture search. Once trained, deploying to a new robot becomes a lightweight search instead of a new training run.
Why one model per device stops scaling
Transformer-based visuomotor policies like DiffusionPolicy, MDT, and OpenVLA now regularly cross hundreds of millions of parameters, and they have to run in the closed loop of a physical robot. Prior compression work in this space, notably SQIL, showed you can get real gains from quantization-aware imitation learning. SQIL reports up to 2.5x speedup and 2.5x energy savings at 4-bit weight quantization on a real edge GPU while preserving accuracy. The catch is that those numbers apply to one model on one device. Change the hardware or the backbone and you retrain.