CARGO: Physical AI for Industrial Package Stacking

TL;DR

CARGO (Contact-Aware Reinforcement-learned Generalized Object-stacking) is Ambi Robotics’ Sim2Real RL policy for high-level task planning of industrial 3D bin-packing — deciding which box to pick next and where to place it, while a separate low-level “AmbiOS” skill stack handles the actual manipulation. Trained across ~40M simulated stack actions on hundreds of NVIDIA GPU nodes with Ray, it hits ≥75% simulated density (>15% over an EMS heuristic baseline) and is already in customer production at 72.5% density and 340+ sorts/hour. The blog positions CARGO as an agentic Physical AI architecture — a reasoning head that composes existing dexterous skills — rather than an end-to-end VLA policy.

Key claims

Industrial online 3D bin-packing is stated as a variant of NP-hard 3D Tetris/knapsack under item-arrival uncertainty; classical solvers assume full item-sequence control, teleop doesn’t scale, and humans struggle at densification — motivating an RL approach [§The opportunity].
CARGO is trained with Sim2Real RL in a proprietary simulator that validates stability + hardware feasibility and returns realistic container states in microseconds, distributed across “hundreds of NVIDIA GPU nodes with Ray” for >40M stack actions [§Scaling].
Scaling simulated placements from ~4M to ~40M during training yields an average +7% density across two container scenarios, and the report frames the experiment as compute-limited rather than saturated [§Scaling, Fig. 1].
On the training distribution CARGO reaches 89% of an offline oracle solver’s density (73.6% vs 82.5%); the gap is attributed to the inherent cost of on-the-fly decisions without knowing future arrivals [§Simulated Packing Performance].
CARGO outperforms an Empty Maximal Space (EMS) online heuristic by 10–15% density on 5 of 6 test distributions, and the paper attributes this to multi-step lookahead + implicit box-distribution priors rather than one-step best-fit [§Simulated Packing Performance].
On unseen distributions with divergent aspect/height ratios, a few hours of post-training on the new distribution recovers the density of training-from-scratch on that distribution — framed as a fast-adaptation property of the base model [Fig. 4].
In live customer production on AmbiStack robots, CARGO achieves >72.5% density at >340 sorts/hour, which the blog claims is comparable-density to humans at 1.5–2× throughput based on the cited literature range [§Introducing CARGO].

Method

CARGO is a high-level action policy — its action space is (which item to pick next, where to place it, in what orientation) — layered on top of AmbiOS’s existing low-level “dexterous manipulation AI skills.” Training is entirely in a proprietary rigid-body simulator that models contact physics, inertial properties, and stack stability under uncertainty, and is engineered specifically for evaluation-speed-per-action (microsecond-scale realistic container states); the blog treats simulator throughput as the load-bearing scaling axis. Rollouts run on hundreds of NVIDIA GPU nodes coordinated by Ray, executing >40M stack actions across randomized item sequences during a single training run.

Two evaluation baselines are used: (a) an anytime multi-start offline oracle solver — a heuristic-guided parallel search over item orderings + orientations with local-search refinement — which upper-bounds any on-the-fly policy, and (b) an Empty Maximal Space (EMS) online heuristic that places each incoming box into the best-fit empty rectangular volume. Post-training on a new box distribution is used as a quick-adaptation mechanism when the aspect/height statistics diverge from pretraining.

Results

Simulated density on training distribution: 73.6% (CARGO) vs 82.5% (oracle) — 89% of oracle [§Simulated Packing Performance].
CARGO beats EMS by 10–15% density on 5 of 6 test distributions; loses only on “Unseen Distribution E” (2 SKUs, sequences constructed for 100% utilization — where the oracle also hits 100% and heuristics are optimal) [§Simulated Packing Performance].
Cross-distribution generalization: 67–72% density on multiple unseen distributions with zero adaptation [§Simulated Packing Performance].
Data-scaling: +7% average density going from ~4M → ~40M training placements, across two container scenarios [Fig. 1].
Live production on AmbiStack: >72.5% density, >340 sorts/hour; framed as ~human-density at 1.5–2× human throughput per [§Introducing CARGO refs 15–20].

Why it’s interesting

CARGO is a rare filed example of Sim2Real RL scaled to a specific, well-defined industrial reasoning task (online 3D packing) and reported as being in real customer production — a design point that sits between end-to-end VLA scaling (VLA Models) and generalist WFM planners (World Foundation Models). Where GEN-1.5 (GEN-1.5: Embodied Foundation Models are One-Shot Learners) and GEN-0 (GEN-0: Embodied Foundation Models That Scale with Physical Interaction) argue for one big embodied foundation model that does everything, CARGO argues the opposite: keep the manipulation skills separate, put the RL only on the task-planning head, and use a fast task-specific simulator rather than a video-based world model as the training substrate — closer in spirit to how Building Worlds That Train Robots — Real-to-Sim-to-Real (R2S2R) as a Scalable Engine for Training and Evaluating Robot Policies frames R2S2R for policy training but restricted to a narrow reasoning subtask. The +7% from 10× simulated actions and the “training was compute-limited, not saturated” framing is a small but useful datapoint for the Reasoning RL scaling picture in a robotics setting.