Flexion Reflect v1.0: The Path Towards Long-Horizon Autonomous Humanoid Work

TL;DR

Flexion’s Reflect v1.0 is a humanoid robotics intelligence platform that composes a fine-tuned VLM mission controller, a motion layer combining a teleop-trained VLA with RL-trained skills, a real-time whole-body controller (Reflex), and a custom low-latency runtime (FlexComm) into a single mission-capable stack. The demo mission — fetch a parcel from another floor via stairs, return via elevator, unpack and shelve the contents from a single natural-language instruction — is end-to-end autonomous with no human teleop, and Flexion reports that RL fine-tuning of the mission VLM lifts a 16-step evaluation from 38% (SFT) to 90% (SFT+RL) end-to-end completion. The headline architectural bet is that RL is used at every layer (low-level control, motion skills, mission reasoning), not just for individual policies, and that semantic global maps with natural-language queries serve as the agent’s spatial memory tool.

Key claims

The mission controller is a fine-tuned VLM that observes the egocentric camera, reasons over progress, and selects the next action through structured tools, with a “semantic map” tool generated from a one-time building scan that supports natural-language area/object queries and global path planning [§2].
Off-the-shelf VLMs fail as mission controllers because they act too eagerly — emitting the next logically plausible tool call before visually verifying the previous call’s completion and the next step’s preconditions [§2].
On a 16-step mission evaluation, the base VLM fails almost immediately, SFT reaches 38% end-to-end completion, and SFT+RL reaches 90% end-to-end completion [§2, RL fine-tuning figure].
Mission specification is purely natural-language — task changes are prompt edits — and the system supports mid-mission interruption where the user issues a new instruction and the agent replans [§2, mission-variation videos].
Below the agent, most motion skills are trained in simulation with custom visual encoders and targeted domain randomization, with RL exposing each policy to a broad distribution of object poses, geometries, and robot start configurations until robust to deployment variation [§3].
A single box-pickup policy generalizes to boxes ranging 100 g to 3.5 kg with varying weight, friction, and deformability; a separate full-body repositioning policy moves the picked box under one arm to free the other [§3, Contact-Rich Whole-Body Interaction].
Dexterous manipulation (unpacking, shelving) is delegated to a VLA trained on teleoperated data with the whole-body controller in the loop, but Flexion explicitly notes this is the brittlest part of the stack and that they are migrating these tasks to RL [§3, Dexterous manipulation].
The low-level whole-body controller (Reflex) runs in real time while respecting balance, actuation limits, and safety constraints; the robot can perform 100+ stair traversals in a row and manipulate under external disturbances without falling [§4].
Recovery happens at two levels: the motion layer retries locally using behaviors learned during RL training (e.g., re-attempting a missed grasp, adjusting when a box is pushed), and the VLM detects off-nominal situations from the camera feed and replans when a subgoal fails [§5].
FlexComm, Flexion’s custom robotics IPC layer, reports tens-to-hundreds-of-microseconds same-host latency, up to 40% speedup and 30% CPU improvement vs. ROS 2 DDS, and resilience to WiFi roaming/disconnects across a deployment spanning many Jetsons, sensors, motors, and servers [§6].
The evaluation/development loop runs the full Reflect pipeline in a photorealistic 3D Gaussian Splatting simulator before real-robot deployment to catch issues earlier [§6, simulation rollout].

Method

Reflect v1.0 is a four-layer stack on a humanoid platform. (1) The mission controller is a custom VLM that consumes the egocentric camera, reasons about mission progress, and emits structured tool calls; tools include a semantic map (precomputed from a building scan, supports natural-language queries and global path requests), navigation, manipulation skills, and elevator/door interaction primitives. (2) The motion layer combines a teleop-data-trained VLA (for dexterous manipulation like unpacking and shelving, with the whole-body controller in the loop) and a library of RL-trained skills (locomotion, contact-rich interaction, box pickup/repositioning, elevator-button pressing, door handling) trained mostly in simulation with custom visual encoders and domain randomization. (3) The low-level whole-body controller Reflex runs in real time and is trained per-robot/morphology with minimal human effort. (4) The runtime FlexComm provides IPC across heterogeneous compute (Jetsons, workstations, cloud), with a monitoring platform that surfaces diagnostics from reasoning traces down to kernel/NIC logs.

The headline training step is RL fine-tuning of the mission VLM. After SFT on demonstration data gets the controller to 38% completion on a 16-step evaluation set (measuring how far each rollout progresses before its first mistake), RL post-training pushes it to 90%. The development loop runs the entire stack in a 3D-Gaussian-Splatting-based simulator that has to satisfy four constraints simultaneously: observations realistic enough for the VLM and perception models, geometry accurate enough for the low- and high-level policies to function, throughput fast enough for realistic loop rates, and full per-decision inspectability for debugging.

Results

Mission VLM reliability: 38% → 90% end-to-end completion on a 16-step mission evaluation (SFT vs SFT+RL); the base off-the-shelf VLM “fails almost immediately” even with careful prompting and in-context examples [§2].
Whole-body controller robustness: 100+ consecutive stair traversals without falling; tracking maintained under external disturbances mid-manipulation [§4].
Manipulation generalization: single box-pickup policy spans 100 g to 3.5 kg payloads with varying friction and deformability [§3].
Runtime: up to 40% speedup and 30% CPU reduction vs ROS 2 DDS for same-host messaging; same-host latencies in the tens-to-hundreds of microseconds [§6].
Demo mission: end-to-end autonomous execution of a multi-floor parcel-retrieval-and-shelving instruction, plus mid-mission re-instruction handled by the same agent [§1, §2].

No external benchmark numbers (no LIBERO, SimplerEnv, RoboArena, UMI-Bench) are reported — evaluation is end-to-end on the company’s own mission distribution, and Flexion explicitly flags this as a limitation: “the current system still operates within a bounded task distribution” [§7].

Why it’s interesting

Reflect v1.0 sits at the integration end of the VLA stack rather than at any single recipe. Most filed VLA work isolates one lever — action pretraining (π*0.6: a VLA That Learns From Experience (RECAP)), clean teleop (Spirit-v1.5: Clean Data Is the Enemy of Great Robot Foundation Models), pointing-grounding (Embodied-R1.5: Evolving Physical Intelligence via Embodied Foundation Models), VLA-as-tool under a VLM orchestrator (VoLo: A Physical Orchestrator for Open-Vocabulary Long-Horizon Manipulation) — and shows it suffices on a benchmark; Reflect explicitly composes a teleop-trained VLA for dexterous manipulation inside an RL-fine-tuned VLM orchestrator and acknowledges that the dexterous-VLA piece is the brittlest layer, with plans to migrate it to RL. This matches VoLo: A Physical Orchestrator for Open-Vocabulary Long-Horizon Manipulation‘s VoLoAgent framing (VLA-as-interruptible-tool) but pushes further into deployment plumbing: FlexComm and the GS-based development simulator are the substrate questions that Inside NVIDIA Halos for Robotics: A Full-Stack Functional Safety System for Physical AI (Halos OS / Safety Island) and LEGS: Fine-Tuning Teleop-Free VLAs for Humanoid Loco-manipulation in an Embodied Gaussian Splatting World (3DGS sim for humanoid VLA fine-tuning) treat separately, integrated here in one stack. The 38%→90% jump from RL fine-tuning of the mission VLM is also the cleanest filed evidence that high-level agent reasoning (not just low-level skills) is the right place to spend RL compute when end-to-end mission completion matters — an empirical anchor for Reasoning RL in an embodied setting that the LLM-only RL papers can’t supply.