π₀.7 and Helix: VLA Transformers Advance Zero-Shot Dexterous Manipulation...

Introduction: The VLA Paradigm Shift in Robotic Intelligence

Vision-Language-Action (VLA) models have emerged as the dominant architecture for embodied AI, unifying pretrained vision-language models (VLMs) with action generation heads to create generalist robotic policies. Building on foundational work like Google's RT-2 (2023, arXiv:2307.15818) which first coined the VLA term and demonstrated 63% gains on novel objects via web-scale token fusion, and UC Berkeley's Octo (2024, RSS, 93M parameters trained on 800k Open X-Embodiment trajectories), recent 2025-2026 releases push toward true autonomy.

Physical Intelligence (π) released π₀ in October 2024 as a flow-matching VLA on VLM backbones, followed by π₀.5 (April 2025) for open-world generalization across unseen kitchens/bedrooms, and π₀.7 (April 2026) emphasizing compositional generalization and steerability. Figure AI's Helix (February 2025) introduced full-upper-body continuous control at 200 Hz on dual Figure robots, marking the first commercial-ready VLA for multi-robot collaboration. NVIDIA's GR00T N1 series (March 2025 onward, arXiv:2503.14734) complements with open humanoid foundation models trained on egocentric videos, simulation, and teleop data.

These advances directly address high-frequency neural feedback loops, zero-shot generalization, and physical torque mapping—core to the 'Brain' pillar. Unlike modular pipelines, VLAs treat actions as sequence predictions, enabling end-to-end optimization from pixels and language to torques.

Architectural Foundations: Transformers, Flow Matching, and Dual Systems

The core software approach in leading VLAs combines transformer encoders for multimodal fusion with specialized action decoders. π₀ employs a flow-matching architecture atop a pretrained VLM (e.g., Gemma-based), inheriting internet-scale semantics while using a DiT-style action expert for continuous control. This contrasts with RT-2's autoregressive DCT/BPE action tokenization and Octo's diffusion decoder.

Helix uses a dual-system design: System 2 (7B-parameter VLM at 7-9 Hz) handles scene understanding and language; System 1 (80M-parameter cross-attention transformer at 200 Hz) maps latents to continuous actions for wrists, torso, head, and fingers. GR00T N1 features a dual-system VLA with FLARE objectives for human ego-video integration.

Recent extensions like ForceVLA (NeurIPS 2025) and FAVLA (arXiv:2602.23648, Feb 2026) introduce force-aware Mixture-of-Experts (MoE) modules, fusing 6-axis force/torque feedback directly into action decoding for contact-rich tasks. This enables high-frequency neural feedback loops—policies operate at 50-500 Hz effective rates via impedance/admittance mapping, far exceeding vision-only baselines.

Zero-shot generalization stems from large-scale pretraining: π₀ on multi-robot, multi-task data; Helix on ~500 hours multi-operator teleop with VLM hindsight labeling; GR00T on 20k+ hours egocentric video scaling laws showing dexterity doubling. π₀.7 demonstrates emergent compositional skills, solving untrained task combinations.

Hardware Integration: Actuators, Sensors, and Torque Mapping

VLA deployment demands tight coupling with high-precision hardware. Dexterous platforms like Shadow Dexterous Hands or Figure humanoids use harmonic drives for backdrivability and low backlash, paired with FOC (Field-Oriented Control) torque-controlled motors. Force/torque sensors (e.g., ATI Mini F/T at wrists) and tactile arrays (GelSight/DIGIT) provide proprioceptive and exteroceptive feedback at kHz rates.

Torque mapping in these systems compares policy outputs to physical dynamics: VLA action chunks (e.g., 10-20 steps) are interpolated to low-level controllers running at 1 kHz. In ForceVLA, FVLMoE dynamically weights visual-language embeddings against real-time wrench measurements, achieving 23%+ success gains on plug insertion (up to 80% vs. π₀ baselines). High-frequency loops close via admittance control, where policies output virtual targets/stiffness rather than raw positions, enabling compliant interaction.

Physical deployments report sub-millimeter precision in peg-in-hole and assembly, with neural observers (e.g., strain-wave gear torque estimation at 5 kHz via NN) compensating for model errors. Compared to older RL baselines (e.g., 2018 DDPG on 24-DoF hands), modern VLAs scale to bimanual long-horizon tasks via cross-embodiment pretraining.

Performance Analysis and Real-World Deployments

Benchmarks show π₀ variants outperforming OpenVLA (7B) and Octo on complex tasks like laundry folding, espresso filter insertion, and bussing—scoring 0.875+ on hard UR5e variants vs. near-zero for baselines. Helix enables novel object pickup and dual-robot collaboration on unseen long-horizon tasks. GR00T N1.7 reports scaling laws: 1k to 20k ego-video hours doubles task completion.

Industry signals include Figure's alpha residential testing (late 2025), NVIDIA partnerships (AeiRobot, NEURA), and Physical Intelligence's RECAP RL fine-tuning doubling throughput. DexVLA (2025) and DexGraspVLA achieve 90%+ zero-shot on unseen grasping via hierarchical planning + diffusion controllers.

Limitations: Latency, Energy, Fragility, and Cost

Current drawbacks persist despite progress. Inference latency remains a bottleneck—large VLMs (7B+) struggle with sub-10ms cycles needed for true 200+ Hz loops without heavy quantization or dual-system splitting. Energy drain is acute on edge GPUs; onboard deployment requires low-power optimizations (Helix claims dual embedded GPUs). Fragility arises in sim-to-real gaps for high-frequency dynamics—policies degrade on novel materials or disturbances without force feedback. Cost barriers include data collection (hundreds of robot-hours per task) and hardware (humanoid platforms exceed $100k). Sample inefficiency lingers; even with pretraining, fine-tuning demands significant real-world interaction.

Unresolved Questions for the Community

Key open problems include: How to scale f